xref: /linux/drivers/gpu/drm/vc4/vc4_validate.c (revision d639d9fa162aadec1ae9980c4dcf6e50bd2f8290)
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 <drm/drm_print.h>
47 
48 #include "vc4_drv.h"
49 #include "vc4_packet.h"
50 
51 #define VALIDATE_ARGS \
52 	struct vc4_exec_info *exec,			\
53 	void *validated,				\
54 	void *untrusted
55 
56 /** Return the width in pixels of a 64-byte microtile. */
57 static uint32_t
58 utile_width(int cpp)
59 {
60 	switch (cpp) {
61 	case 1:
62 	case 2:
63 		return 8;
64 	case 4:
65 		return 4;
66 	case 8:
67 		return 2;
68 	default:
69 		pr_err("unknown cpp: %d\n", cpp);
70 		return 1;
71 	}
72 }
73 
74 /** Return the height in pixels of a 64-byte microtile. */
75 static uint32_t
76 utile_height(int cpp)
77 {
78 	switch (cpp) {
79 	case 1:
80 		return 8;
81 	case 2:
82 	case 4:
83 	case 8:
84 		return 4;
85 	default:
86 		pr_err("unknown cpp: %d\n", cpp);
87 		return 1;
88 	}
89 }
90 
91 /**
92  * size_is_lt() - Returns whether a miplevel of the given size will
93  * use the lineartile (LT) tiling layout rather than the normal T
94  * tiling layout.
95  * @width: Width in pixels of the miplevel
96  * @height: Height in pixels of the miplevel
97  * @cpp: Bytes per pixel of the pixel format
98  */
99 static bool
100 size_is_lt(uint32_t width, uint32_t height, int cpp)
101 {
102 	return (width <= 4 * utile_width(cpp) ||
103 		height <= 4 * utile_height(cpp));
104 }
105 
106 struct drm_gem_dma_object *
107 vc4_use_bo(struct vc4_exec_info *exec, uint32_t hindex)
108 {
109 	struct vc4_dev *vc4 = exec->dev;
110 	struct drm_gem_dma_object *obj;
111 	struct vc4_bo *bo;
112 
113 	if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
114 		return NULL;
115 
116 	if (hindex >= exec->bo_count) {
117 		DRM_DEBUG("BO index %d greater than BO count %d\n",
118 			  hindex, exec->bo_count);
119 		return NULL;
120 	}
121 	obj = to_drm_gem_dma_obj(exec->bo[hindex]);
122 	bo = to_vc4_bo(&obj->base);
123 
124 	if (bo->validated_shader) {
125 		DRM_DEBUG("Trying to use shader BO as something other than "
126 			  "a shader\n");
127 		return NULL;
128 	}
129 
130 	return obj;
131 }
132 
133 static struct drm_gem_dma_object *
134 vc4_use_handle(struct vc4_exec_info *exec, uint32_t gem_handles_packet_index)
135 {
136 	return vc4_use_bo(exec, exec->bo_index[gem_handles_packet_index]);
137 }
138 
139 static bool
140 validate_bin_pos(struct vc4_exec_info *exec, void *untrusted, uint32_t pos)
141 {
142 	/* Note that the untrusted pointer passed to these functions is
143 	 * incremented past the packet byte.
144 	 */
145 	return (untrusted - 1 == exec->bin_u + pos);
146 }
147 
148 static uint32_t
149 gl_shader_rec_size(uint32_t pointer_bits)
150 {
151 	uint32_t attribute_count = pointer_bits & 7;
152 	bool extended = pointer_bits & 8;
153 
154 	if (attribute_count == 0)
155 		attribute_count = 8;
156 
157 	if (extended)
158 		return 100 + attribute_count * 4;
159 	else
160 		return 36 + attribute_count * 8;
161 }
162 
163 bool
164 vc4_check_tex_size(struct vc4_exec_info *exec, struct drm_gem_dma_object *fbo,
165 		   uint32_t offset, uint8_t tiling_format,
166 		   uint32_t width, uint32_t height, uint8_t cpp)
167 {
168 	struct vc4_dev *vc4 = exec->dev;
169 	uint32_t aligned_width, aligned_height, stride, size;
170 	uint32_t utile_w = utile_width(cpp);
171 	uint32_t utile_h = utile_height(cpp);
172 
173 	if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
174 		return false;
175 
176 	/* The shaded vertex format stores signed 12.4 fixed point
177 	 * (-2048,2047) offsets from the viewport center, so we should
178 	 * never have a render target larger than 4096.  The texture
179 	 * unit can only sample from 2048x2048, so it's even more
180 	 * restricted.  This lets us avoid worrying about overflow in
181 	 * our math.
182 	 */
183 	if (width > 4096 || height > 4096) {
184 		DRM_DEBUG("Surface dimensions (%d,%d) too large",
185 			  width, height);
186 		return false;
187 	}
188 
189 	switch (tiling_format) {
190 	case VC4_TILING_FORMAT_LINEAR:
191 		aligned_width = round_up(width, utile_w);
192 		aligned_height = height;
193 		break;
194 	case VC4_TILING_FORMAT_T:
195 		aligned_width = round_up(width, utile_w * 8);
196 		aligned_height = round_up(height, utile_h * 8);
197 		break;
198 	case VC4_TILING_FORMAT_LT:
199 		aligned_width = round_up(width, utile_w);
200 		aligned_height = round_up(height, utile_h);
201 		break;
202 	default:
203 		DRM_DEBUG("buffer tiling %d unsupported\n", tiling_format);
204 		return false;
205 	}
206 
207 	stride = aligned_width * cpp;
208 	size = stride * aligned_height;
209 
210 	if (size + offset < size ||
211 	    size + offset > fbo->base.size) {
212 		DRM_DEBUG("Overflow in %dx%d (%dx%d) fbo size (%d + %d > %zd)\n",
213 			  width, height,
214 			  aligned_width, aligned_height,
215 			  size, offset, fbo->base.size);
216 		return false;
217 	}
218 
219 	return true;
220 }
221 
222 static int
223 validate_flush(VALIDATE_ARGS)
224 {
225 	if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 1)) {
226 		DRM_DEBUG("Bin CL must end with VC4_PACKET_FLUSH\n");
227 		return -EINVAL;
228 	}
229 	exec->found_flush = true;
230 
231 	return 0;
232 }
233 
234 static int
235 validate_start_tile_binning(VALIDATE_ARGS)
236 {
237 	if (exec->found_start_tile_binning_packet) {
238 		DRM_DEBUG("Duplicate VC4_PACKET_START_TILE_BINNING\n");
239 		return -EINVAL;
240 	}
241 	exec->found_start_tile_binning_packet = true;
242 
243 	if (!exec->found_tile_binning_mode_config_packet) {
244 		DRM_DEBUG("missing VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
245 		return -EINVAL;
246 	}
247 
248 	return 0;
249 }
250 
251 static int
252 validate_increment_semaphore(VALIDATE_ARGS)
253 {
254 	if (!validate_bin_pos(exec, untrusted, exec->args->bin_cl_size - 2)) {
255 		DRM_DEBUG("Bin CL must end with "
256 			  "VC4_PACKET_INCREMENT_SEMAPHORE\n");
257 		return -EINVAL;
258 	}
259 	exec->found_increment_semaphore_packet = true;
260 
261 	return 0;
262 }
263 
264 static int
265 validate_indexed_prim_list(VALIDATE_ARGS)
266 {
267 	struct drm_gem_dma_object *ib;
268 	uint32_t length = *(uint32_t *)(untrusted + 1);
269 	uint32_t offset = *(uint32_t *)(untrusted + 5);
270 	uint32_t max_index = *(uint32_t *)(untrusted + 9);
271 	uint32_t index_size = (*(uint8_t *)(untrusted + 0) >> 4) ? 2 : 1;
272 	struct vc4_shader_state *shader_state;
273 
274 	/* Check overflow condition */
275 	if (exec->shader_state_count == 0) {
276 		DRM_DEBUG("shader state must precede primitives\n");
277 		return -EINVAL;
278 	}
279 	shader_state = &exec->shader_state[exec->shader_state_count - 1];
280 
281 	if (max_index > shader_state->max_index)
282 		shader_state->max_index = max_index;
283 
284 	ib = vc4_use_handle(exec, 0);
285 	if (!ib)
286 		return -EINVAL;
287 
288 	if (offset > ib->base.size ||
289 	    (ib->base.size - offset) / index_size < length) {
290 		DRM_DEBUG("IB access overflow (%d + %d*%d > %zd)\n",
291 			  offset, length, index_size, ib->base.size);
292 		return -EINVAL;
293 	}
294 
295 	*(uint32_t *)(validated + 5) = ib->dma_addr + offset;
296 
297 	return 0;
298 }
299 
300 static int
301 validate_gl_array_primitive(VALIDATE_ARGS)
302 {
303 	uint32_t length = *(uint32_t *)(untrusted + 1);
304 	uint32_t base_index = *(uint32_t *)(untrusted + 5);
305 	uint32_t max_index;
306 	struct vc4_shader_state *shader_state;
307 
308 	/* Check overflow condition */
309 	if (exec->shader_state_count == 0) {
310 		DRM_DEBUG("shader state must precede primitives\n");
311 		return -EINVAL;
312 	}
313 	shader_state = &exec->shader_state[exec->shader_state_count - 1];
314 
315 	if (length + base_index < length) {
316 		DRM_DEBUG("primitive vertex count overflow\n");
317 		return -EINVAL;
318 	}
319 	max_index = length + base_index - 1;
320 
321 	if (max_index > shader_state->max_index)
322 		shader_state->max_index = max_index;
323 
324 	return 0;
325 }
326 
327 static int
328 validate_gl_shader_state(VALIDATE_ARGS)
329 {
330 	uint32_t i = exec->shader_state_count++;
331 
332 	if (i >= exec->shader_state_size) {
333 		DRM_DEBUG("More requests for shader states than declared\n");
334 		return -EINVAL;
335 	}
336 
337 	exec->shader_state[i].addr = *(uint32_t *)untrusted;
338 	exec->shader_state[i].max_index = 0;
339 
340 	if (exec->shader_state[i].addr & ~0xf) {
341 		DRM_DEBUG("high bits set in GL shader rec reference\n");
342 		return -EINVAL;
343 	}
344 
345 	*(uint32_t *)validated = (exec->shader_rec_p +
346 				  exec->shader_state[i].addr);
347 
348 	exec->shader_rec_p +=
349 		roundup(gl_shader_rec_size(exec->shader_state[i].addr), 16);
350 
351 	return 0;
352 }
353 
354 static int
355 validate_tile_binning_config(VALIDATE_ARGS)
356 {
357 	struct drm_device *dev = exec->exec_bo->base.dev;
358 	struct vc4_dev *vc4 = to_vc4_dev(dev);
359 	uint8_t flags;
360 	uint32_t tile_state_size;
361 	uint32_t tile_count, bin_addr;
362 	int bin_slot;
363 
364 	if (exec->found_tile_binning_mode_config_packet) {
365 		DRM_DEBUG("Duplicate VC4_PACKET_TILE_BINNING_MODE_CONFIG\n");
366 		return -EINVAL;
367 	}
368 	exec->found_tile_binning_mode_config_packet = true;
369 
370 	exec->bin_tiles_x = *(uint8_t *)(untrusted + 12);
371 	exec->bin_tiles_y = *(uint8_t *)(untrusted + 13);
372 	tile_count = exec->bin_tiles_x * exec->bin_tiles_y;
373 	flags = *(uint8_t *)(untrusted + 14);
374 
375 	if (exec->bin_tiles_x == 0 ||
376 	    exec->bin_tiles_y == 0) {
377 		DRM_DEBUG("Tile binning config of %dx%d too small\n",
378 			  exec->bin_tiles_x, exec->bin_tiles_y);
379 		return -EINVAL;
380 	}
381 
382 	if (flags & (VC4_BIN_CONFIG_DB_NON_MS |
383 		     VC4_BIN_CONFIG_TILE_BUFFER_64BIT)) {
384 		DRM_DEBUG("unsupported binning config flags 0x%02x\n", flags);
385 		return -EINVAL;
386 	}
387 
388 	bin_slot = vc4_v3d_get_bin_slot(vc4);
389 	if (bin_slot < 0) {
390 		if (bin_slot != -EINTR && bin_slot != -ERESTARTSYS) {
391 			drm_err(dev, "Failed to allocate binner memory: %d\n",
392 				bin_slot);
393 		}
394 		return bin_slot;
395 	}
396 
397 	/* The slot we allocated will only be used by this job, and is
398 	 * free when the job completes rendering.
399 	 */
400 	exec->bin_slots |= BIT(bin_slot);
401 	bin_addr = vc4->bin_bo->base.dma_addr + bin_slot * vc4->bin_alloc_size;
402 
403 	/* The tile state data array is 48 bytes per tile, and we put it at
404 	 * the start of a BO containing both it and the tile alloc.
405 	 */
406 	tile_state_size = 48 * tile_count;
407 
408 	/* Since the tile alloc array will follow us, align. */
409 	exec->tile_alloc_offset = bin_addr + roundup(tile_state_size, 4096);
410 
411 	*(uint8_t *)(validated + 14) =
412 		((flags & ~(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_MASK |
413 			    VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_MASK)) |
414 		 VC4_BIN_CONFIG_AUTO_INIT_TSDA |
415 		 VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE_32,
416 			       VC4_BIN_CONFIG_ALLOC_INIT_BLOCK_SIZE) |
417 		 VC4_SET_FIELD(VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE_128,
418 			       VC4_BIN_CONFIG_ALLOC_BLOCK_SIZE));
419 
420 	/* tile alloc address. */
421 	*(uint32_t *)(validated + 0) = exec->tile_alloc_offset;
422 	/* tile alloc size. */
423 	*(uint32_t *)(validated + 4) = (bin_addr + vc4->bin_alloc_size -
424 					exec->tile_alloc_offset);
425 	/* tile state address. */
426 	*(uint32_t *)(validated + 8) = bin_addr;
427 
428 	return 0;
429 }
430 
431 static int
432 validate_gem_handles(VALIDATE_ARGS)
433 {
434 	memcpy(exec->bo_index, untrusted, sizeof(exec->bo_index));
435 	return 0;
436 }
437 
438 #define VC4_DEFINE_PACKET(packet, func) \
439 	[packet] = { packet ## _SIZE, #packet, func }
440 
441 static const struct cmd_info {
442 	uint16_t len;
443 	const char *name;
444 	int (*func)(struct vc4_exec_info *exec, void *validated,
445 		    void *untrusted);
446 } cmd_info[] = {
447 	VC4_DEFINE_PACKET(VC4_PACKET_HALT, NULL),
448 	VC4_DEFINE_PACKET(VC4_PACKET_NOP, NULL),
449 	VC4_DEFINE_PACKET(VC4_PACKET_FLUSH, validate_flush),
450 	VC4_DEFINE_PACKET(VC4_PACKET_FLUSH_ALL, NULL),
451 	VC4_DEFINE_PACKET(VC4_PACKET_START_TILE_BINNING,
452 			  validate_start_tile_binning),
453 	VC4_DEFINE_PACKET(VC4_PACKET_INCREMENT_SEMAPHORE,
454 			  validate_increment_semaphore),
455 
456 	VC4_DEFINE_PACKET(VC4_PACKET_GL_INDEXED_PRIMITIVE,
457 			  validate_indexed_prim_list),
458 	VC4_DEFINE_PACKET(VC4_PACKET_GL_ARRAY_PRIMITIVE,
459 			  validate_gl_array_primitive),
460 
461 	VC4_DEFINE_PACKET(VC4_PACKET_PRIMITIVE_LIST_FORMAT, NULL),
462 
463 	VC4_DEFINE_PACKET(VC4_PACKET_GL_SHADER_STATE, validate_gl_shader_state),
464 
465 	VC4_DEFINE_PACKET(VC4_PACKET_CONFIGURATION_BITS, NULL),
466 	VC4_DEFINE_PACKET(VC4_PACKET_FLAT_SHADE_FLAGS, NULL),
467 	VC4_DEFINE_PACKET(VC4_PACKET_POINT_SIZE, NULL),
468 	VC4_DEFINE_PACKET(VC4_PACKET_LINE_WIDTH, NULL),
469 	VC4_DEFINE_PACKET(VC4_PACKET_RHT_X_BOUNDARY, NULL),
470 	VC4_DEFINE_PACKET(VC4_PACKET_DEPTH_OFFSET, NULL),
471 	VC4_DEFINE_PACKET(VC4_PACKET_CLIP_WINDOW, NULL),
472 	VC4_DEFINE_PACKET(VC4_PACKET_VIEWPORT_OFFSET, NULL),
473 	VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_XY_SCALING, NULL),
474 	/* Note: The docs say this was also 105, but it was 106 in the
475 	 * initial userland code drop.
476 	 */
477 	VC4_DEFINE_PACKET(VC4_PACKET_CLIPPER_Z_SCALING, NULL),
478 
479 	VC4_DEFINE_PACKET(VC4_PACKET_TILE_BINNING_MODE_CONFIG,
480 			  validate_tile_binning_config),
481 
482 	VC4_DEFINE_PACKET(VC4_PACKET_GEM_HANDLES, validate_gem_handles),
483 };
484 
485 int
486 vc4_validate_bin_cl(struct drm_device *dev,
487 		    void *validated,
488 		    void *unvalidated,
489 		    struct vc4_exec_info *exec)
490 {
491 	struct vc4_dev *vc4 = to_vc4_dev(dev);
492 	uint32_t len = exec->args->bin_cl_size;
493 	uint32_t dst_offset = 0;
494 	uint32_t src_offset = 0;
495 
496 	if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
497 		return -ENODEV;
498 
499 	while (src_offset < len) {
500 		void *dst_pkt = validated + dst_offset;
501 		void *src_pkt = unvalidated + src_offset;
502 		u8 cmd = *(uint8_t *)src_pkt;
503 		const struct cmd_info *info;
504 
505 		if (cmd >= ARRAY_SIZE(cmd_info)) {
506 			DRM_DEBUG("0x%08x: packet %d out of bounds\n",
507 				  src_offset, cmd);
508 			return -EINVAL;
509 		}
510 
511 		info = &cmd_info[cmd];
512 		if (!info->name) {
513 			DRM_DEBUG("0x%08x: packet %d invalid\n",
514 				  src_offset, cmd);
515 			return -EINVAL;
516 		}
517 
518 		if (src_offset + info->len > len) {
519 			DRM_DEBUG("0x%08x: packet %d (%s) length 0x%08x "
520 				  "exceeds bounds (0x%08x)\n",
521 				  src_offset, cmd, info->name, info->len,
522 				  src_offset + len);
523 			return -EINVAL;
524 		}
525 
526 		if (cmd != VC4_PACKET_GEM_HANDLES)
527 			memcpy(dst_pkt, src_pkt, info->len);
528 
529 		if (info->func && info->func(exec,
530 					     dst_pkt + 1,
531 					     src_pkt + 1)) {
532 			DRM_DEBUG("0x%08x: packet %d (%s) failed to validate\n",
533 				  src_offset, cmd, info->name);
534 			return -EINVAL;
535 		}
536 
537 		src_offset += info->len;
538 		/* GEM handle loading doesn't produce HW packets. */
539 		if (cmd != VC4_PACKET_GEM_HANDLES)
540 			dst_offset += info->len;
541 
542 		/* When the CL hits halt, it'll stop reading anything else. */
543 		if (cmd == VC4_PACKET_HALT)
544 			break;
545 	}
546 
547 	exec->ct0ea = exec->ct0ca + dst_offset;
548 
549 	if (!exec->found_start_tile_binning_packet) {
550 		DRM_DEBUG("Bin CL missing VC4_PACKET_START_TILE_BINNING\n");
551 		return -EINVAL;
552 	}
553 
554 	/* The bin CL must be ended with INCREMENT_SEMAPHORE and FLUSH.  The
555 	 * semaphore is used to trigger the render CL to start up, and the
556 	 * FLUSH is what caps the bin lists with
557 	 * VC4_PACKET_RETURN_FROM_SUB_LIST (so they jump back to the main
558 	 * render CL when they get called to) and actually triggers the queued
559 	 * semaphore increment.
560 	 */
561 	if (!exec->found_increment_semaphore_packet || !exec->found_flush) {
562 		DRM_DEBUG("Bin CL missing VC4_PACKET_INCREMENT_SEMAPHORE + "
563 			  "VC4_PACKET_FLUSH\n");
564 		return -EINVAL;
565 	}
566 
567 	return 0;
568 }
569 
570 static bool
571 reloc_tex(struct vc4_exec_info *exec,
572 	  void *uniform_data_u,
573 	  struct vc4_texture_sample_info *sample,
574 	  uint32_t texture_handle_index, bool is_cs)
575 {
576 	struct drm_gem_dma_object *tex;
577 	uint32_t p0 = *(uint32_t *)(uniform_data_u + sample->p_offset[0]);
578 	uint32_t p1 = *(uint32_t *)(uniform_data_u + sample->p_offset[1]);
579 	uint32_t p2 = (sample->p_offset[2] != ~0 ?
580 		       *(uint32_t *)(uniform_data_u + sample->p_offset[2]) : 0);
581 	uint32_t p3 = (sample->p_offset[3] != ~0 ?
582 		       *(uint32_t *)(uniform_data_u + sample->p_offset[3]) : 0);
583 	uint32_t *validated_p0 = exec->uniforms_v + sample->p_offset[0];
584 	uint32_t offset = p0 & VC4_TEX_P0_OFFSET_MASK;
585 	uint32_t miplevels = VC4_GET_FIELD(p0, VC4_TEX_P0_MIPLVLS);
586 	uint32_t width = VC4_GET_FIELD(p1, VC4_TEX_P1_WIDTH);
587 	uint32_t height = VC4_GET_FIELD(p1, VC4_TEX_P1_HEIGHT);
588 	uint32_t cpp, tiling_format, utile_w, utile_h;
589 	uint32_t i;
590 	uint32_t cube_map_stride = 0;
591 	enum vc4_texture_data_type type;
592 
593 	tex = vc4_use_bo(exec, texture_handle_index);
594 	if (!tex)
595 		return false;
596 
597 	if (sample->is_direct) {
598 		uint32_t remaining_size = tex->base.size - p0;
599 
600 		if (p0 > tex->base.size - 4) {
601 			DRM_DEBUG("UBO offset greater than UBO size\n");
602 			goto fail;
603 		}
604 		if (p1 > remaining_size - 4) {
605 			DRM_DEBUG("UBO clamp would allow reads "
606 				  "outside of UBO\n");
607 			goto fail;
608 		}
609 		*validated_p0 = tex->dma_addr + p0;
610 		return true;
611 	}
612 
613 	if (width == 0)
614 		width = 2048;
615 	if (height == 0)
616 		height = 2048;
617 
618 	if (p0 & VC4_TEX_P0_CMMODE_MASK) {
619 		if (VC4_GET_FIELD(p2, VC4_TEX_P2_PTYPE) ==
620 		    VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE)
621 			cube_map_stride = p2 & VC4_TEX_P2_CMST_MASK;
622 		if (VC4_GET_FIELD(p3, VC4_TEX_P2_PTYPE) ==
623 		    VC4_TEX_P2_PTYPE_CUBE_MAP_STRIDE) {
624 			if (cube_map_stride) {
625 				DRM_DEBUG("Cube map stride set twice\n");
626 				goto fail;
627 			}
628 
629 			cube_map_stride = p3 & VC4_TEX_P2_CMST_MASK;
630 		}
631 		if (!cube_map_stride) {
632 			DRM_DEBUG("Cube map stride not set\n");
633 			goto fail;
634 		}
635 	}
636 
637 	type = (VC4_GET_FIELD(p0, VC4_TEX_P0_TYPE) |
638 		(VC4_GET_FIELD(p1, VC4_TEX_P1_TYPE4) << 4));
639 
640 	switch (type) {
641 	case VC4_TEXTURE_TYPE_RGBA8888:
642 	case VC4_TEXTURE_TYPE_RGBX8888:
643 	case VC4_TEXTURE_TYPE_RGBA32R:
644 		cpp = 4;
645 		break;
646 	case VC4_TEXTURE_TYPE_RGBA4444:
647 	case VC4_TEXTURE_TYPE_RGBA5551:
648 	case VC4_TEXTURE_TYPE_RGB565:
649 	case VC4_TEXTURE_TYPE_LUMALPHA:
650 	case VC4_TEXTURE_TYPE_S16F:
651 	case VC4_TEXTURE_TYPE_S16:
652 		cpp = 2;
653 		break;
654 	case VC4_TEXTURE_TYPE_LUMINANCE:
655 	case VC4_TEXTURE_TYPE_ALPHA:
656 	case VC4_TEXTURE_TYPE_S8:
657 		cpp = 1;
658 		break;
659 	case VC4_TEXTURE_TYPE_ETC1:
660 		/* ETC1 is arranged as 64-bit blocks, where each block is 4x4
661 		 * pixels.
662 		 */
663 		cpp = 8;
664 		width = (width + 3) >> 2;
665 		height = (height + 3) >> 2;
666 		break;
667 	case VC4_TEXTURE_TYPE_BW1:
668 	case VC4_TEXTURE_TYPE_A4:
669 	case VC4_TEXTURE_TYPE_A1:
670 	case VC4_TEXTURE_TYPE_RGBA64:
671 	case VC4_TEXTURE_TYPE_YUV422R:
672 	default:
673 		DRM_DEBUG("Texture format %d unsupported\n", type);
674 		goto fail;
675 	}
676 	utile_w = utile_width(cpp);
677 	utile_h = utile_height(cpp);
678 
679 	if (type == VC4_TEXTURE_TYPE_RGBA32R) {
680 		tiling_format = VC4_TILING_FORMAT_LINEAR;
681 	} else {
682 		if (size_is_lt(width, height, cpp))
683 			tiling_format = VC4_TILING_FORMAT_LT;
684 		else
685 			tiling_format = VC4_TILING_FORMAT_T;
686 	}
687 
688 	if (!vc4_check_tex_size(exec, tex, offset + cube_map_stride * 5,
689 				tiling_format, width, height, cpp)) {
690 		goto fail;
691 	}
692 
693 	/* The mipmap levels are stored before the base of the texture.  Make
694 	 * sure there is actually space in the BO.
695 	 */
696 	for (i = 1; i <= miplevels; i++) {
697 		uint32_t level_width = max(width >> i, 1u);
698 		uint32_t level_height = max(height >> i, 1u);
699 		uint32_t aligned_width, aligned_height;
700 		uint32_t level_size;
701 
702 		/* Once the levels get small enough, they drop from T to LT. */
703 		if (tiling_format == VC4_TILING_FORMAT_T &&
704 		    size_is_lt(level_width, level_height, cpp)) {
705 			tiling_format = VC4_TILING_FORMAT_LT;
706 		}
707 
708 		switch (tiling_format) {
709 		case VC4_TILING_FORMAT_T:
710 			aligned_width = round_up(level_width, utile_w * 8);
711 			aligned_height = round_up(level_height, utile_h * 8);
712 			break;
713 		case VC4_TILING_FORMAT_LT:
714 			aligned_width = round_up(level_width, utile_w);
715 			aligned_height = round_up(level_height, utile_h);
716 			break;
717 		default:
718 			aligned_width = round_up(level_width, utile_w);
719 			aligned_height = level_height;
720 			break;
721 		}
722 
723 		level_size = aligned_width * cpp * aligned_height;
724 
725 		if (offset < level_size) {
726 			DRM_DEBUG("Level %d (%dx%d -> %dx%d) size %db "
727 				  "overflowed buffer bounds (offset %d)\n",
728 				  i, level_width, level_height,
729 				  aligned_width, aligned_height,
730 				  level_size, offset);
731 			goto fail;
732 		}
733 
734 		offset -= level_size;
735 	}
736 
737 	*validated_p0 = tex->dma_addr + p0;
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_dma_object *bo[ARRAY_SIZE(shader_reloc_offsets) + 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_DEBUG("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_DEBUG("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_DEBUG("Shader handle %d too big\n", src_handles[i]);
803 			return -EINVAL;
804 		}
805 
806 		bo[i] = to_drm_gem_dma_obj(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_DEBUG("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_DEBUG("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]->dma_addr + src_offset;
837 
838 		if (src_offset != 0) {
839 			DRM_DEBUG("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_DEBUG("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_dma_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 		if (state->addr & 0x8)
901 			stride |= (*(uint32_t *)(pkt_u + 100 + i * 4)) & ~0xff;
902 
903 		if (vbo->base.size < offset ||
904 		    vbo->base.size - offset < attr_size) {
905 			DRM_DEBUG("BO offset overflow (%d + %d > %zu)\n",
906 				  offset, attr_size, vbo->base.size);
907 			return -EINVAL;
908 		}
909 
910 		if (stride != 0) {
911 			max_index = ((vbo->base.size - offset - attr_size) /
912 				     stride);
913 			if (state->max_index > max_index) {
914 				DRM_DEBUG("primitives use index %d out of "
915 					  "supplied %d\n",
916 					  state->max_index, max_index);
917 				return -EINVAL;
918 			}
919 		}
920 
921 		*(uint32_t *)(pkt_v + o) = vbo->dma_addr + offset;
922 	}
923 
924 	return 0;
925 }
926 
927 int
928 vc4_validate_shader_recs(struct drm_device *dev,
929 			 struct vc4_exec_info *exec)
930 {
931 	struct vc4_dev *vc4 = to_vc4_dev(dev);
932 	uint32_t i;
933 	int ret = 0;
934 
935 	if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
936 		return -ENODEV;
937 
938 	for (i = 0; i < exec->shader_state_count; i++) {
939 		ret = validate_gl_shader_rec(dev, exec, &exec->shader_state[i]);
940 		if (ret)
941 			return ret;
942 	}
943 
944 	return ret;
945 }
946