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: Shader validator for VC4.
26  *
27  * The VC4 has no IOMMU between it and system memory, so a user with
28  * access to execute shaders could escalate privilege by overwriting
29  * system memory (using the VPM write address register in the
30  * general-purpose DMA mode) or reading system memory it shouldn't
31  * (reading it as a texture, or uniform data, or vertex data).
32  *
33  * This walks over a shader BO, ensuring that its accesses are
34  * appropriately bounded, and recording how many texture accesses are
35  * made and where so that we can do relocations for them in the
36  * uniform stream.
37  */
38 
39 #include "vc4_drv.h"
40 #include "vc4_qpu_defines.h"
41 
42 #define LIVE_REG_COUNT (32 + 32 + 4)
43 
44 struct vc4_shader_validation_state {
45 	/* Current IP being validated. */
46 	uint32_t ip;
47 
48 	/* IP at the end of the BO, do not read shader[max_ip] */
49 	uint32_t max_ip;
50 
51 	uint64_t *shader;
52 
53 	struct vc4_texture_sample_info tmu_setup[2];
54 	int tmu_write_count[2];
55 
56 	/* For registers that were last written to by a MIN instruction with
57 	 * one argument being a uniform, the address of the uniform.
58 	 * Otherwise, ~0.
59 	 *
60 	 * This is used for the validation of direct address memory reads.
61 	 */
62 	uint32_t live_min_clamp_offsets[LIVE_REG_COUNT];
63 	bool live_max_clamp_regs[LIVE_REG_COUNT];
64 	uint32_t live_immediates[LIVE_REG_COUNT];
65 
66 	/* Bitfield of which IPs are used as branch targets.
67 	 *
68 	 * Used for validation that the uniform stream is updated at the right
69 	 * points and clearing the texturing/clamping state.
70 	 */
71 	unsigned long *branch_targets;
72 
73 	/* Set when entering a basic block, and cleared when the uniform
74 	 * address update is found.  This is used to make sure that we don't
75 	 * read uniforms when the address is undefined.
76 	 */
77 	bool needs_uniform_address_update;
78 
79 	/* Set when we find a backwards branch.  If the branch is backwards,
80 	 * the taraget is probably doing an address reset to read uniforms,
81 	 * and so we need to be sure that a uniforms address is present in the
82 	 * stream, even if the shader didn't need to read uniforms in later
83 	 * basic blocks.
84 	 */
85 	bool needs_uniform_address_for_loop;
86 };
87 
88 static uint32_t
89 waddr_to_live_reg_index(uint32_t waddr, bool is_b)
90 {
91 	if (waddr < 32) {
92 		if (is_b)
93 			return 32 + waddr;
94 		else
95 			return waddr;
96 	} else if (waddr <= QPU_W_ACC3) {
97 		return 64 + waddr - QPU_W_ACC0;
98 	} else {
99 		return ~0;
100 	}
101 }
102 
103 static uint32_t
104 raddr_add_a_to_live_reg_index(uint64_t inst)
105 {
106 	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
107 	uint32_t add_a = QPU_GET_FIELD(inst, QPU_ADD_A);
108 	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
109 	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
110 
111 	if (add_a == QPU_MUX_A)
112 		return raddr_a;
113 	else if (add_a == QPU_MUX_B && sig != QPU_SIG_SMALL_IMM)
114 		return 32 + raddr_b;
115 	else if (add_a <= QPU_MUX_R3)
116 		return 64 + add_a;
117 	else
118 		return ~0;
119 }
120 
121 static bool
122 is_tmu_submit(uint32_t waddr)
123 {
124 	return (waddr == QPU_W_TMU0_S ||
125 		waddr == QPU_W_TMU1_S);
126 }
127 
128 static bool
129 is_tmu_write(uint32_t waddr)
130 {
131 	return (waddr >= QPU_W_TMU0_S &&
132 		waddr <= QPU_W_TMU1_B);
133 }
134 
135 static bool
136 record_texture_sample(struct vc4_validated_shader_info *validated_shader,
137 		      struct vc4_shader_validation_state *validation_state,
138 		      int tmu)
139 {
140 	uint32_t s = validated_shader->num_texture_samples;
141 	int i;
142 	struct vc4_texture_sample_info *temp_samples;
143 
144 	temp_samples = krealloc(validated_shader->texture_samples,
145 				(s + 1) * sizeof(*temp_samples),
146 				GFP_KERNEL);
147 	if (!temp_samples)
148 		return false;
149 
150 	memcpy(&temp_samples[s],
151 	       &validation_state->tmu_setup[tmu],
152 	       sizeof(*temp_samples));
153 
154 	validated_shader->num_texture_samples = s + 1;
155 	validated_shader->texture_samples = temp_samples;
156 
157 	for (i = 0; i < 4; i++)
158 		validation_state->tmu_setup[tmu].p_offset[i] = ~0;
159 
160 	return true;
161 }
162 
163 static bool
164 check_tmu_write(struct vc4_validated_shader_info *validated_shader,
165 		struct vc4_shader_validation_state *validation_state,
166 		bool is_mul)
167 {
168 	uint64_t inst = validation_state->shader[validation_state->ip];
169 	uint32_t waddr = (is_mul ?
170 			  QPU_GET_FIELD(inst, QPU_WADDR_MUL) :
171 			  QPU_GET_FIELD(inst, QPU_WADDR_ADD));
172 	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
173 	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
174 	int tmu = waddr > QPU_W_TMU0_B;
175 	bool submit = is_tmu_submit(waddr);
176 	bool is_direct = submit && validation_state->tmu_write_count[tmu] == 0;
177 	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
178 
179 	if (is_direct) {
180 		uint32_t add_b = QPU_GET_FIELD(inst, QPU_ADD_B);
181 		uint32_t clamp_reg, clamp_offset;
182 
183 		if (sig == QPU_SIG_SMALL_IMM) {
184 			DRM_ERROR("direct TMU read used small immediate\n");
185 			return false;
186 		}
187 
188 		/* Make sure that this texture load is an add of the base
189 		 * address of the UBO to a clamped offset within the UBO.
190 		 */
191 		if (is_mul ||
192 		    QPU_GET_FIELD(inst, QPU_OP_ADD) != QPU_A_ADD) {
193 			DRM_ERROR("direct TMU load wasn't an add\n");
194 			return false;
195 		}
196 
197 		/* We assert that the clamped address is the first
198 		 * argument, and the UBO base address is the second argument.
199 		 * This is arbitrary, but simpler than supporting flipping the
200 		 * two either way.
201 		 */
202 		clamp_reg = raddr_add_a_to_live_reg_index(inst);
203 		if (clamp_reg == ~0) {
204 			DRM_ERROR("direct TMU load wasn't clamped\n");
205 			return false;
206 		}
207 
208 		clamp_offset = validation_state->live_min_clamp_offsets[clamp_reg];
209 		if (clamp_offset == ~0) {
210 			DRM_ERROR("direct TMU load wasn't clamped\n");
211 			return false;
212 		}
213 
214 		/* Store the clamp value's offset in p1 (see reloc_tex() in
215 		 * vc4_validate.c).
216 		 */
217 		validation_state->tmu_setup[tmu].p_offset[1] =
218 			clamp_offset;
219 
220 		if (!(add_b == QPU_MUX_A && raddr_a == QPU_R_UNIF) &&
221 		    !(add_b == QPU_MUX_B && raddr_b == QPU_R_UNIF)) {
222 			DRM_ERROR("direct TMU load didn't add to a uniform\n");
223 			return false;
224 		}
225 
226 		validation_state->tmu_setup[tmu].is_direct = true;
227 	} else {
228 		if (raddr_a == QPU_R_UNIF || (sig != QPU_SIG_SMALL_IMM &&
229 					      raddr_b == QPU_R_UNIF)) {
230 			DRM_ERROR("uniform read in the same instruction as "
231 				  "texture setup.\n");
232 			return false;
233 		}
234 	}
235 
236 	if (validation_state->tmu_write_count[tmu] >= 4) {
237 		DRM_ERROR("TMU%d got too many parameters before dispatch\n",
238 			  tmu);
239 		return false;
240 	}
241 	validation_state->tmu_setup[tmu].p_offset[validation_state->tmu_write_count[tmu]] =
242 		validated_shader->uniforms_size;
243 	validation_state->tmu_write_count[tmu]++;
244 	/* Since direct uses a RADDR uniform reference, it will get counted in
245 	 * check_instruction_reads()
246 	 */
247 	if (!is_direct) {
248 		if (validation_state->needs_uniform_address_update) {
249 			DRM_ERROR("Texturing with undefined uniform address\n");
250 			return false;
251 		}
252 
253 		validated_shader->uniforms_size += 4;
254 	}
255 
256 	if (submit) {
257 		if (!record_texture_sample(validated_shader,
258 					   validation_state, tmu)) {
259 			return false;
260 		}
261 
262 		validation_state->tmu_write_count[tmu] = 0;
263 	}
264 
265 	return true;
266 }
267 
268 static bool require_uniform_address_uniform(struct vc4_validated_shader_info *validated_shader)
269 {
270 	uint32_t o = validated_shader->num_uniform_addr_offsets;
271 	uint32_t num_uniforms = validated_shader->uniforms_size / 4;
272 
273 	validated_shader->uniform_addr_offsets =
274 		krealloc(validated_shader->uniform_addr_offsets,
275 			 (o + 1) *
276 			 sizeof(*validated_shader->uniform_addr_offsets),
277 			 GFP_KERNEL);
278 	if (!validated_shader->uniform_addr_offsets)
279 		return false;
280 
281 	validated_shader->uniform_addr_offsets[o] = num_uniforms;
282 	validated_shader->num_uniform_addr_offsets++;
283 
284 	return true;
285 }
286 
287 static bool
288 validate_uniform_address_write(struct vc4_validated_shader_info *validated_shader,
289 			       struct vc4_shader_validation_state *validation_state,
290 			       bool is_mul)
291 {
292 	uint64_t inst = validation_state->shader[validation_state->ip];
293 	u32 add_b = QPU_GET_FIELD(inst, QPU_ADD_B);
294 	u32 raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
295 	u32 raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
296 	u32 add_lri = raddr_add_a_to_live_reg_index(inst);
297 	/* We want our reset to be pointing at whatever uniform follows the
298 	 * uniforms base address.
299 	 */
300 	u32 expected_offset = validated_shader->uniforms_size + 4;
301 
302 	/* We only support absolute uniform address changes, and we
303 	 * require that they be in the current basic block before any
304 	 * of its uniform reads.
305 	 *
306 	 * One could potentially emit more efficient QPU code, by
307 	 * noticing that (say) an if statement does uniform control
308 	 * flow for all threads and that the if reads the same number
309 	 * of uniforms on each side.  However, this scheme is easy to
310 	 * validate so it's all we allow for now.
311 	 */
312 	switch (QPU_GET_FIELD(inst, QPU_SIG)) {
313 	case QPU_SIG_NONE:
314 	case QPU_SIG_SCOREBOARD_UNLOCK:
315 	case QPU_SIG_COLOR_LOAD:
316 	case QPU_SIG_LOAD_TMU0:
317 	case QPU_SIG_LOAD_TMU1:
318 		break;
319 	default:
320 		DRM_ERROR("uniforms address change must be "
321 			  "normal math\n");
322 		return false;
323 	}
324 
325 	if (is_mul || QPU_GET_FIELD(inst, QPU_OP_ADD) != QPU_A_ADD) {
326 		DRM_ERROR("Uniform address reset must be an ADD.\n");
327 		return false;
328 	}
329 
330 	if (QPU_GET_FIELD(inst, QPU_COND_ADD) != QPU_COND_ALWAYS) {
331 		DRM_ERROR("Uniform address reset must be unconditional.\n");
332 		return false;
333 	}
334 
335 	if (QPU_GET_FIELD(inst, QPU_PACK) != QPU_PACK_A_NOP &&
336 	    !(inst & QPU_PM)) {
337 		DRM_ERROR("No packing allowed on uniforms reset\n");
338 		return false;
339 	}
340 
341 	if (add_lri == -1) {
342 		DRM_ERROR("First argument of uniform address write must be "
343 			  "an immediate value.\n");
344 		return false;
345 	}
346 
347 	if (validation_state->live_immediates[add_lri] != expected_offset) {
348 		DRM_ERROR("Resetting uniforms with offset %db instead of %db\n",
349 			  validation_state->live_immediates[add_lri],
350 			  expected_offset);
351 		return false;
352 	}
353 
354 	if (!(add_b == QPU_MUX_A && raddr_a == QPU_R_UNIF) &&
355 	    !(add_b == QPU_MUX_B && raddr_b == QPU_R_UNIF)) {
356 		DRM_ERROR("Second argument of uniform address write must be "
357 			  "a uniform.\n");
358 		return false;
359 	}
360 
361 	validation_state->needs_uniform_address_update = false;
362 	validation_state->needs_uniform_address_for_loop = false;
363 	return require_uniform_address_uniform(validated_shader);
364 }
365 
366 static bool
367 check_reg_write(struct vc4_validated_shader_info *validated_shader,
368 		struct vc4_shader_validation_state *validation_state,
369 		bool is_mul)
370 {
371 	uint64_t inst = validation_state->shader[validation_state->ip];
372 	uint32_t waddr = (is_mul ?
373 			  QPU_GET_FIELD(inst, QPU_WADDR_MUL) :
374 			  QPU_GET_FIELD(inst, QPU_WADDR_ADD));
375 	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
376 	bool ws = inst & QPU_WS;
377 	bool is_b = is_mul ^ ws;
378 	u32 lri = waddr_to_live_reg_index(waddr, is_b);
379 
380 	if (lri != -1) {
381 		uint32_t cond_add = QPU_GET_FIELD(inst, QPU_COND_ADD);
382 		uint32_t cond_mul = QPU_GET_FIELD(inst, QPU_COND_MUL);
383 
384 		if (sig == QPU_SIG_LOAD_IMM &&
385 		    QPU_GET_FIELD(inst, QPU_PACK) == QPU_PACK_A_NOP &&
386 		    ((is_mul && cond_mul == QPU_COND_ALWAYS) ||
387 		     (!is_mul && cond_add == QPU_COND_ALWAYS))) {
388 			validation_state->live_immediates[lri] =
389 				QPU_GET_FIELD(inst, QPU_LOAD_IMM);
390 		} else {
391 			validation_state->live_immediates[lri] = ~0;
392 		}
393 	}
394 
395 	switch (waddr) {
396 	case QPU_W_UNIFORMS_ADDRESS:
397 		if (is_b) {
398 			DRM_ERROR("relative uniforms address change "
399 				  "unsupported\n");
400 			return false;
401 		}
402 
403 		return validate_uniform_address_write(validated_shader,
404 						      validation_state,
405 						      is_mul);
406 
407 	case QPU_W_TLB_COLOR_MS:
408 	case QPU_W_TLB_COLOR_ALL:
409 	case QPU_W_TLB_Z:
410 		/* These only interact with the tile buffer, not main memory,
411 		 * so they're safe.
412 		 */
413 		return true;
414 
415 	case QPU_W_TMU0_S:
416 	case QPU_W_TMU0_T:
417 	case QPU_W_TMU0_R:
418 	case QPU_W_TMU0_B:
419 	case QPU_W_TMU1_S:
420 	case QPU_W_TMU1_T:
421 	case QPU_W_TMU1_R:
422 	case QPU_W_TMU1_B:
423 		return check_tmu_write(validated_shader, validation_state,
424 				       is_mul);
425 
426 	case QPU_W_HOST_INT:
427 	case QPU_W_TMU_NOSWAP:
428 	case QPU_W_TLB_ALPHA_MASK:
429 	case QPU_W_MUTEX_RELEASE:
430 		/* XXX: I haven't thought about these, so don't support them
431 		 * for now.
432 		 */
433 		DRM_ERROR("Unsupported waddr %d\n", waddr);
434 		return false;
435 
436 	case QPU_W_VPM_ADDR:
437 		DRM_ERROR("General VPM DMA unsupported\n");
438 		return false;
439 
440 	case QPU_W_VPM:
441 	case QPU_W_VPMVCD_SETUP:
442 		/* We allow VPM setup in general, even including VPM DMA
443 		 * configuration setup, because the (unsafe) DMA can only be
444 		 * triggered by QPU_W_VPM_ADDR writes.
445 		 */
446 		return true;
447 
448 	case QPU_W_TLB_STENCIL_SETUP:
449 		return true;
450 	}
451 
452 	return true;
453 }
454 
455 static void
456 track_live_clamps(struct vc4_validated_shader_info *validated_shader,
457 		  struct vc4_shader_validation_state *validation_state)
458 {
459 	uint64_t inst = validation_state->shader[validation_state->ip];
460 	uint32_t op_add = QPU_GET_FIELD(inst, QPU_OP_ADD);
461 	uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
462 	uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
463 	uint32_t cond_add = QPU_GET_FIELD(inst, QPU_COND_ADD);
464 	uint32_t add_a = QPU_GET_FIELD(inst, QPU_ADD_A);
465 	uint32_t add_b = QPU_GET_FIELD(inst, QPU_ADD_B);
466 	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
467 	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
468 	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
469 	bool ws = inst & QPU_WS;
470 	uint32_t lri_add_a, lri_add, lri_mul;
471 	bool add_a_is_min_0;
472 
473 	/* Check whether OP_ADD's A argumennt comes from a live MAX(x, 0),
474 	 * before we clear previous live state.
475 	 */
476 	lri_add_a = raddr_add_a_to_live_reg_index(inst);
477 	add_a_is_min_0 = (lri_add_a != ~0 &&
478 			  validation_state->live_max_clamp_regs[lri_add_a]);
479 
480 	/* Clear live state for registers written by our instruction. */
481 	lri_add = waddr_to_live_reg_index(waddr_add, ws);
482 	lri_mul = waddr_to_live_reg_index(waddr_mul, !ws);
483 	if (lri_mul != ~0) {
484 		validation_state->live_max_clamp_regs[lri_mul] = false;
485 		validation_state->live_min_clamp_offsets[lri_mul] = ~0;
486 	}
487 	if (lri_add != ~0) {
488 		validation_state->live_max_clamp_regs[lri_add] = false;
489 		validation_state->live_min_clamp_offsets[lri_add] = ~0;
490 	} else {
491 		/* Nothing further to do for live tracking, since only ADDs
492 		 * generate new live clamp registers.
493 		 */
494 		return;
495 	}
496 
497 	/* Now, handle remaining live clamp tracking for the ADD operation. */
498 
499 	if (cond_add != QPU_COND_ALWAYS)
500 		return;
501 
502 	if (op_add == QPU_A_MAX) {
503 		/* Track live clamps of a value to a minimum of 0 (in either
504 		 * arg).
505 		 */
506 		if (sig != QPU_SIG_SMALL_IMM || raddr_b != 0 ||
507 		    (add_a != QPU_MUX_B && add_b != QPU_MUX_B)) {
508 			return;
509 		}
510 
511 		validation_state->live_max_clamp_regs[lri_add] = true;
512 	} else if (op_add == QPU_A_MIN) {
513 		/* Track live clamps of a value clamped to a minimum of 0 and
514 		 * a maximum of some uniform's offset.
515 		 */
516 		if (!add_a_is_min_0)
517 			return;
518 
519 		if (!(add_b == QPU_MUX_A && raddr_a == QPU_R_UNIF) &&
520 		    !(add_b == QPU_MUX_B && raddr_b == QPU_R_UNIF &&
521 		      sig != QPU_SIG_SMALL_IMM)) {
522 			return;
523 		}
524 
525 		validation_state->live_min_clamp_offsets[lri_add] =
526 			validated_shader->uniforms_size;
527 	}
528 }
529 
530 static bool
531 check_instruction_writes(struct vc4_validated_shader_info *validated_shader,
532 			 struct vc4_shader_validation_state *validation_state)
533 {
534 	uint64_t inst = validation_state->shader[validation_state->ip];
535 	uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
536 	uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
537 	bool ok;
538 
539 	if (is_tmu_write(waddr_add) && is_tmu_write(waddr_mul)) {
540 		DRM_ERROR("ADD and MUL both set up textures\n");
541 		return false;
542 	}
543 
544 	ok = (check_reg_write(validated_shader, validation_state, false) &&
545 	      check_reg_write(validated_shader, validation_state, true));
546 
547 	track_live_clamps(validated_shader, validation_state);
548 
549 	return ok;
550 }
551 
552 static bool
553 check_branch(uint64_t inst,
554 	     struct vc4_validated_shader_info *validated_shader,
555 	     struct vc4_shader_validation_state *validation_state,
556 	     int ip)
557 {
558 	int32_t branch_imm = QPU_GET_FIELD(inst, QPU_BRANCH_TARGET);
559 	uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
560 	uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
561 
562 	if ((int)branch_imm < 0)
563 		validation_state->needs_uniform_address_for_loop = true;
564 
565 	/* We don't want to have to worry about validation of this, and
566 	 * there's no need for it.
567 	 */
568 	if (waddr_add != QPU_W_NOP || waddr_mul != QPU_W_NOP) {
569 		DRM_ERROR("branch instruction at %d wrote a register.\n",
570 			  validation_state->ip);
571 		return false;
572 	}
573 
574 	return true;
575 }
576 
577 static bool
578 check_instruction_reads(struct vc4_validated_shader_info *validated_shader,
579 			struct vc4_shader_validation_state *validation_state)
580 {
581 	uint64_t inst = validation_state->shader[validation_state->ip];
582 	uint32_t raddr_a = QPU_GET_FIELD(inst, QPU_RADDR_A);
583 	uint32_t raddr_b = QPU_GET_FIELD(inst, QPU_RADDR_B);
584 	uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
585 
586 	if (raddr_a == QPU_R_UNIF ||
587 	    (raddr_b == QPU_R_UNIF && sig != QPU_SIG_SMALL_IMM)) {
588 		/* This can't overflow the uint32_t, because we're reading 8
589 		 * bytes of instruction to increment by 4 here, so we'd
590 		 * already be OOM.
591 		 */
592 		validated_shader->uniforms_size += 4;
593 
594 		if (validation_state->needs_uniform_address_update) {
595 			DRM_ERROR("Uniform read with undefined uniform "
596 				  "address\n");
597 			return false;
598 		}
599 	}
600 
601 	return true;
602 }
603 
604 /* Make sure that all branches are absolute and point within the shader, and
605  * note their targets for later.
606  */
607 static bool
608 vc4_validate_branches(struct vc4_shader_validation_state *validation_state)
609 {
610 	uint32_t max_branch_target = 0;
611 	bool found_shader_end = false;
612 	int ip;
613 	int shader_end_ip = 0;
614 	int last_branch = -2;
615 
616 	for (ip = 0; ip < validation_state->max_ip; ip++) {
617 		uint64_t inst = validation_state->shader[ip];
618 		int32_t branch_imm = QPU_GET_FIELD(inst, QPU_BRANCH_TARGET);
619 		uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
620 		uint32_t after_delay_ip = ip + 4;
621 		uint32_t branch_target_ip;
622 
623 		if (sig == QPU_SIG_PROG_END) {
624 			shader_end_ip = ip;
625 			found_shader_end = true;
626 			continue;
627 		}
628 
629 		if (sig != QPU_SIG_BRANCH)
630 			continue;
631 
632 		if (ip - last_branch < 4) {
633 			DRM_ERROR("Branch at %d during delay slots\n", ip);
634 			return false;
635 		}
636 		last_branch = ip;
637 
638 		if (inst & QPU_BRANCH_REG) {
639 			DRM_ERROR("branching from register relative "
640 				  "not supported\n");
641 			return false;
642 		}
643 
644 		if (!(inst & QPU_BRANCH_REL)) {
645 			DRM_ERROR("relative branching required\n");
646 			return false;
647 		}
648 
649 		/* The actual branch target is the instruction after the delay
650 		 * slots, plus whatever byte offset is in the low 32 bits of
651 		 * the instruction.  Make sure we're not branching beyond the
652 		 * end of the shader object.
653 		 */
654 		if (branch_imm % sizeof(inst) != 0) {
655 			DRM_ERROR("branch target not aligned\n");
656 			return false;
657 		}
658 
659 		branch_target_ip = after_delay_ip + (branch_imm >> 3);
660 		if (branch_target_ip >= validation_state->max_ip) {
661 			DRM_ERROR("Branch at %d outside of shader (ip %d/%d)\n",
662 				  ip, branch_target_ip,
663 				  validation_state->max_ip);
664 			return false;
665 		}
666 		set_bit(branch_target_ip, validation_state->branch_targets);
667 
668 		/* Make sure that the non-branching path is also not outside
669 		 * the shader.
670 		 */
671 		if (after_delay_ip >= validation_state->max_ip) {
672 			DRM_ERROR("Branch at %d continues past shader end "
673 				  "(%d/%d)\n",
674 				  ip, after_delay_ip, validation_state->max_ip);
675 			return false;
676 		}
677 		set_bit(after_delay_ip, validation_state->branch_targets);
678 		max_branch_target = max(max_branch_target, after_delay_ip);
679 
680 		/* There are two delay slots after program end is signaled
681 		 * that are still executed, then we're finished.
682 		 */
683 		if (found_shader_end && ip == shader_end_ip + 2)
684 			break;
685 	}
686 
687 	if (max_branch_target > shader_end_ip) {
688 		DRM_ERROR("Branch landed after QPU_SIG_PROG_END");
689 		return false;
690 	}
691 
692 	return true;
693 }
694 
695 /* Resets any known state for the shader, used when we may be branched to from
696  * multiple locations in the program (or at shader start).
697  */
698 static void
699 reset_validation_state(struct vc4_shader_validation_state *validation_state)
700 {
701 	int i;
702 
703 	for (i = 0; i < 8; i++)
704 		validation_state->tmu_setup[i / 4].p_offset[i % 4] = ~0;
705 
706 	for (i = 0; i < LIVE_REG_COUNT; i++) {
707 		validation_state->live_min_clamp_offsets[i] = ~0;
708 		validation_state->live_max_clamp_regs[i] = false;
709 		validation_state->live_immediates[i] = ~0;
710 	}
711 }
712 
713 static bool
714 texturing_in_progress(struct vc4_shader_validation_state *validation_state)
715 {
716 	return (validation_state->tmu_write_count[0] != 0 ||
717 		validation_state->tmu_write_count[1] != 0);
718 }
719 
720 static bool
721 vc4_handle_branch_target(struct vc4_shader_validation_state *validation_state)
722 {
723 	uint32_t ip = validation_state->ip;
724 
725 	if (!test_bit(ip, validation_state->branch_targets))
726 		return true;
727 
728 	if (texturing_in_progress(validation_state)) {
729 		DRM_ERROR("Branch target landed during TMU setup\n");
730 		return false;
731 	}
732 
733 	/* Reset our live values tracking, since this instruction may have
734 	 * multiple predecessors.
735 	 *
736 	 * One could potentially do analysis to determine that, for
737 	 * example, all predecessors have a live max clamp in the same
738 	 * register, but we don't bother with that.
739 	 */
740 	reset_validation_state(validation_state);
741 
742 	/* Since we've entered a basic block from potentially multiple
743 	 * predecessors, we need the uniforms address to be updated before any
744 	 * unforms are read.  We require that after any branch point, the next
745 	 * uniform to be loaded is a uniform address offset.  That uniform's
746 	 * offset will be marked by the uniform address register write
747 	 * validation, or a one-off the end-of-program check.
748 	 */
749 	validation_state->needs_uniform_address_update = true;
750 
751 	return true;
752 }
753 
754 struct vc4_validated_shader_info *
755 vc4_validate_shader(struct drm_gem_cma_object *shader_obj)
756 {
757 	bool found_shader_end = false;
758 	int shader_end_ip = 0;
759 	uint32_t ip;
760 	struct vc4_validated_shader_info *validated_shader = NULL;
761 	struct vc4_shader_validation_state validation_state;
762 
763 	memset(&validation_state, 0, sizeof(validation_state));
764 	validation_state.shader = shader_obj->vaddr;
765 	validation_state.max_ip = shader_obj->base.size / sizeof(uint64_t);
766 
767 	reset_validation_state(&validation_state);
768 
769 	validation_state.branch_targets =
770 		kcalloc(BITS_TO_LONGS(validation_state.max_ip),
771 			sizeof(unsigned long), GFP_KERNEL);
772 	if (!validation_state.branch_targets)
773 		goto fail;
774 
775 	validated_shader = kcalloc(1, sizeof(*validated_shader), GFP_KERNEL);
776 	if (!validated_shader)
777 		goto fail;
778 
779 	if (!vc4_validate_branches(&validation_state))
780 		goto fail;
781 
782 	for (ip = 0; ip < validation_state.max_ip; ip++) {
783 		uint64_t inst = validation_state.shader[ip];
784 		uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
785 
786 		validation_state.ip = ip;
787 
788 		if (!vc4_handle_branch_target(&validation_state))
789 			goto fail;
790 
791 		switch (sig) {
792 		case QPU_SIG_NONE:
793 		case QPU_SIG_WAIT_FOR_SCOREBOARD:
794 		case QPU_SIG_SCOREBOARD_UNLOCK:
795 		case QPU_SIG_COLOR_LOAD:
796 		case QPU_SIG_LOAD_TMU0:
797 		case QPU_SIG_LOAD_TMU1:
798 		case QPU_SIG_PROG_END:
799 		case QPU_SIG_SMALL_IMM:
800 			if (!check_instruction_writes(validated_shader,
801 						      &validation_state)) {
802 				DRM_ERROR("Bad write at ip %d\n", ip);
803 				goto fail;
804 			}
805 
806 			if (!check_instruction_reads(validated_shader,
807 						     &validation_state))
808 				goto fail;
809 
810 			if (sig == QPU_SIG_PROG_END) {
811 				found_shader_end = true;
812 				shader_end_ip = ip;
813 			}
814 
815 			break;
816 
817 		case QPU_SIG_LOAD_IMM:
818 			if (!check_instruction_writes(validated_shader,
819 						      &validation_state)) {
820 				DRM_ERROR("Bad LOAD_IMM write at ip %d\n", ip);
821 				goto fail;
822 			}
823 			break;
824 
825 		case QPU_SIG_BRANCH:
826 			if (!check_branch(inst, validated_shader,
827 					  &validation_state, ip))
828 				goto fail;
829 			break;
830 		default:
831 			DRM_ERROR("Unsupported QPU signal %d at "
832 				  "instruction %d\n", sig, ip);
833 			goto fail;
834 		}
835 
836 		/* There are two delay slots after program end is signaled
837 		 * that are still executed, then we're finished.
838 		 */
839 		if (found_shader_end && ip == shader_end_ip + 2)
840 			break;
841 	}
842 
843 	if (ip == validation_state.max_ip) {
844 		DRM_ERROR("shader failed to terminate before "
845 			  "shader BO end at %zd\n",
846 			  shader_obj->base.size);
847 		goto fail;
848 	}
849 
850 	/* If we did a backwards branch and we haven't emitted a uniforms
851 	 * reset since then, we still need the uniforms stream to have the
852 	 * uniforms address available so that the backwards branch can do its
853 	 * uniforms reset.
854 	 *
855 	 * We could potentially prove that the backwards branch doesn't
856 	 * contain any uses of uniforms until program exit, but that doesn't
857 	 * seem to be worth the trouble.
858 	 */
859 	if (validation_state.needs_uniform_address_for_loop) {
860 		if (!require_uniform_address_uniform(validated_shader))
861 			goto fail;
862 		validated_shader->uniforms_size += 4;
863 	}
864 
865 	/* Again, no chance of integer overflow here because the worst case
866 	 * scenario is 8 bytes of uniforms plus handles per 8-byte
867 	 * instruction.
868 	 */
869 	validated_shader->uniforms_src_size =
870 		(validated_shader->uniforms_size +
871 		 4 * validated_shader->num_texture_samples);
872 
873 	kfree(validation_state.branch_targets);
874 
875 	return validated_shader;
876 
877 fail:
878 	kfree(validation_state.branch_targets);
879 	if (validated_shader) {
880 		kfree(validated_shader->texture_samples);
881 		kfree(validated_shader);
882 	}
883 	return NULL;
884 }
885