2 * Copyright © 2014 Broadcom
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:
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
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
25 * DOC: Shader validator for VC4.
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).
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
41 #include "vc4_qpu_defines.h"
43 #define LIVE_REG_COUNT (32 + 32 + 4)
45 struct vc4_shader_validation_state
{
46 /* Current IP being validated. */
49 /* IP at the end of the BO, do not read shader[max_ip] */
54 struct vc4_texture_sample_info tmu_setup
[2];
55 int tmu_write_count
[2];
57 /* For registers that were last written to by a MIN instruction with
58 * one argument being a uniform, the address of the uniform.
61 * This is used for the validation of direct address memory reads.
63 uint32_t live_min_clamp_offsets
[LIVE_REG_COUNT
];
64 bool live_max_clamp_regs
[LIVE_REG_COUNT
];
65 uint32_t live_immediates
[LIVE_REG_COUNT
];
67 /* Bitfield of which IPs are used as branch targets.
69 * Used for validation that the uniform stream is updated at the right
70 * points and clearing the texturing/clamping state.
72 unsigned long *branch_targets
;
74 /* Set when entering a basic block, and cleared when the uniform
75 * address update is found. This is used to make sure that we don't
76 * read uniforms when the address is undefined.
78 bool needs_uniform_address_update
;
80 /* Set when we find a backwards branch. If the branch is backwards,
81 * the taraget is probably doing an address reset to read uniforms,
82 * and so we need to be sure that a uniforms address is present in the
83 * stream, even if the shader didn't need to read uniforms in later
86 bool needs_uniform_address_for_loop
;
90 waddr_to_live_reg_index(uint32_t waddr
, bool is_b
)
97 } else if (waddr
<= QPU_W_ACC3
) {
98 return 64 + waddr
- QPU_W_ACC0
;
105 raddr_add_a_to_live_reg_index(uint64_t inst
)
107 uint32_t sig
= QPU_GET_FIELD(inst
, QPU_SIG
);
108 uint32_t add_a
= QPU_GET_FIELD(inst
, QPU_ADD_A
);
109 uint32_t raddr_a
= QPU_GET_FIELD(inst
, QPU_RADDR_A
);
110 uint32_t raddr_b
= QPU_GET_FIELD(inst
, QPU_RADDR_B
);
112 if (add_a
== QPU_MUX_A
)
114 else if (add_a
== QPU_MUX_B
&& sig
!= QPU_SIG_SMALL_IMM
)
116 else if (add_a
<= QPU_MUX_R3
)
123 is_tmu_submit(uint32_t waddr
)
125 return (waddr
== QPU_W_TMU0_S
||
126 waddr
== QPU_W_TMU1_S
);
130 is_tmu_write(uint32_t waddr
)
132 return (waddr
>= QPU_W_TMU0_S
&&
133 waddr
<= QPU_W_TMU1_B
);
137 record_texture_sample(struct vc4_validated_shader_info
*validated_shader
,
138 struct vc4_shader_validation_state
*validation_state
,
141 uint32_t s
= validated_shader
->num_texture_samples
;
143 struct vc4_texture_sample_info
*temp_samples
;
145 temp_samples
= krealloc(validated_shader
->texture_samples
,
146 (s
+ 1) * sizeof(*temp_samples
),
151 memcpy(&temp_samples
[s
],
152 &validation_state
->tmu_setup
[tmu
],
153 sizeof(*temp_samples
));
155 validated_shader
->num_texture_samples
= s
+ 1;
156 validated_shader
->texture_samples
= temp_samples
;
158 for (i
= 0; i
< 4; i
++)
159 validation_state
->tmu_setup
[tmu
].p_offset
[i
] = ~0;
165 check_tmu_write(struct vc4_validated_shader_info
*validated_shader
,
166 struct vc4_shader_validation_state
*validation_state
,
169 uint64_t inst
= validation_state
->shader
[validation_state
->ip
];
170 uint32_t waddr
= (is_mul
?
171 QPU_GET_FIELD(inst
, QPU_WADDR_MUL
) :
172 QPU_GET_FIELD(inst
, QPU_WADDR_ADD
));
173 uint32_t raddr_a
= QPU_GET_FIELD(inst
, QPU_RADDR_A
);
174 uint32_t raddr_b
= QPU_GET_FIELD(inst
, QPU_RADDR_B
);
175 int tmu
= waddr
> QPU_W_TMU0_B
;
176 bool submit
= is_tmu_submit(waddr
);
177 bool is_direct
= submit
&& validation_state
->tmu_write_count
[tmu
] == 0;
178 uint32_t sig
= QPU_GET_FIELD(inst
, QPU_SIG
);
181 uint32_t add_b
= QPU_GET_FIELD(inst
, QPU_ADD_B
);
182 uint32_t clamp_reg
, clamp_offset
;
184 if (sig
== QPU_SIG_SMALL_IMM
) {
185 DRM_ERROR("direct TMU read used small immediate\n");
189 /* Make sure that this texture load is an add of the base
190 * address of the UBO to a clamped offset within the UBO.
193 QPU_GET_FIELD(inst
, QPU_OP_ADD
) != QPU_A_ADD
) {
194 DRM_ERROR("direct TMU load wasn't an add\n");
198 /* We assert that the clamped address is the first
199 * argument, and the UBO base address is the second argument.
200 * This is arbitrary, but simpler than supporting flipping the
203 clamp_reg
= raddr_add_a_to_live_reg_index(inst
);
204 if (clamp_reg
== ~0) {
205 DRM_ERROR("direct TMU load wasn't clamped\n");
209 clamp_offset
= validation_state
->live_min_clamp_offsets
[clamp_reg
];
210 if (clamp_offset
== ~0) {
211 DRM_ERROR("direct TMU load wasn't clamped\n");
215 /* Store the clamp value's offset in p1 (see reloc_tex() in
218 validation_state
->tmu_setup
[tmu
].p_offset
[1] =
221 if (!(add_b
== QPU_MUX_A
&& raddr_a
== QPU_R_UNIF
) &&
222 !(add_b
== QPU_MUX_B
&& raddr_b
== QPU_R_UNIF
)) {
223 DRM_ERROR("direct TMU load didn't add to a uniform\n");
227 validation_state
->tmu_setup
[tmu
].is_direct
= true;
229 if (raddr_a
== QPU_R_UNIF
|| (sig
!= QPU_SIG_SMALL_IMM
&&
230 raddr_b
== QPU_R_UNIF
)) {
231 DRM_ERROR("uniform read in the same instruction as "
237 if (validation_state
->tmu_write_count
[tmu
] >= 4) {
238 DRM_ERROR("TMU%d got too many parameters before dispatch\n",
242 validation_state
->tmu_setup
[tmu
].p_offset
[validation_state
->tmu_write_count
[tmu
]] =
243 validated_shader
->uniforms_size
;
244 validation_state
->tmu_write_count
[tmu
]++;
245 /* Since direct uses a RADDR uniform reference, it will get counted in
246 * check_instruction_reads()
249 if (validation_state
->needs_uniform_address_update
) {
250 DRM_ERROR("Texturing with undefined uniform address\n");
254 validated_shader
->uniforms_size
+= 4;
258 if (!record_texture_sample(validated_shader
,
259 validation_state
, tmu
)) {
263 validation_state
->tmu_write_count
[tmu
] = 0;
269 static bool require_uniform_address_uniform(struct vc4_validated_shader_info
*validated_shader
)
271 uint32_t o
= validated_shader
->num_uniform_addr_offsets
;
272 uint32_t num_uniforms
= validated_shader
->uniforms_size
/ 4;
274 validated_shader
->uniform_addr_offsets
=
275 krealloc(validated_shader
->uniform_addr_offsets
,
277 sizeof(*validated_shader
->uniform_addr_offsets
),
279 if (!validated_shader
->uniform_addr_offsets
)
282 validated_shader
->uniform_addr_offsets
[o
] = num_uniforms
;
283 validated_shader
->num_uniform_addr_offsets
++;
289 validate_uniform_address_write(struct vc4_validated_shader_info
*validated_shader
,
290 struct vc4_shader_validation_state
*validation_state
,
293 uint64_t inst
= validation_state
->shader
[validation_state
->ip
];
294 u32 add_b
= QPU_GET_FIELD(inst
, QPU_ADD_B
);
295 u32 raddr_a
= QPU_GET_FIELD(inst
, QPU_RADDR_A
);
296 u32 raddr_b
= QPU_GET_FIELD(inst
, QPU_RADDR_B
);
297 u32 add_lri
= raddr_add_a_to_live_reg_index(inst
);
298 /* We want our reset to be pointing at whatever uniform follows the
299 * uniforms base address.
301 u32 expected_offset
= validated_shader
->uniforms_size
+ 4;
303 /* We only support absolute uniform address changes, and we
304 * require that they be in the current basic block before any
305 * of its uniform reads.
307 * One could potentially emit more efficient QPU code, by
308 * noticing that (say) an if statement does uniform control
309 * flow for all threads and that the if reads the same number
310 * of uniforms on each side. However, this scheme is easy to
311 * validate so it's all we allow for now.
314 if (QPU_GET_FIELD(inst
, QPU_SIG
) != QPU_SIG_NONE
) {
315 DRM_ERROR("uniforms address change must be "
320 if (is_mul
|| QPU_GET_FIELD(inst
, QPU_OP_ADD
) != QPU_A_ADD
) {
321 DRM_ERROR("Uniform address reset must be an ADD.\n");
325 if (QPU_GET_FIELD(inst
, QPU_COND_ADD
) != QPU_COND_ALWAYS
) {
326 DRM_ERROR("Uniform address reset must be unconditional.\n");
330 if (QPU_GET_FIELD(inst
, QPU_PACK
) != QPU_PACK_A_NOP
&&
332 DRM_ERROR("No packing allowed on uniforms reset\n");
337 DRM_ERROR("First argument of uniform address write must be "
338 "an immediate value.\n");
342 if (validation_state
->live_immediates
[add_lri
] != expected_offset
) {
343 DRM_ERROR("Resetting uniforms with offset %db instead of %db\n",
344 validation_state
->live_immediates
[add_lri
],
349 if (!(add_b
== QPU_MUX_A
&& raddr_a
== QPU_R_UNIF
) &&
350 !(add_b
== QPU_MUX_B
&& raddr_b
== QPU_R_UNIF
)) {
351 DRM_ERROR("Second argument of uniform address write must be "
356 validation_state
->needs_uniform_address_update
= false;
357 validation_state
->needs_uniform_address_for_loop
= false;
358 return require_uniform_address_uniform(validated_shader
);
362 check_reg_write(struct vc4_validated_shader_info
*validated_shader
,
363 struct vc4_shader_validation_state
*validation_state
,
366 uint64_t inst
= validation_state
->shader
[validation_state
->ip
];
367 uint32_t waddr
= (is_mul
?
368 QPU_GET_FIELD(inst
, QPU_WADDR_MUL
) :
369 QPU_GET_FIELD(inst
, QPU_WADDR_ADD
));
370 uint32_t sig
= QPU_GET_FIELD(inst
, QPU_SIG
);
371 bool ws
= inst
& QPU_WS
;
372 bool is_b
= is_mul
^ ws
;
373 u32 lri
= waddr_to_live_reg_index(waddr
, is_b
);
376 uint32_t cond_add
= QPU_GET_FIELD(inst
, QPU_COND_ADD
);
377 uint32_t cond_mul
= QPU_GET_FIELD(inst
, QPU_COND_MUL
);
379 if (sig
== QPU_SIG_LOAD_IMM
&&
380 QPU_GET_FIELD(inst
, QPU_PACK
) == QPU_PACK_A_NOP
&&
381 ((is_mul
&& cond_mul
== QPU_COND_ALWAYS
) ||
382 (!is_mul
&& cond_add
== QPU_COND_ALWAYS
))) {
383 validation_state
->live_immediates
[lri
] =
384 QPU_GET_FIELD(inst
, QPU_LOAD_IMM
);
386 validation_state
->live_immediates
[lri
] = ~0;
391 case QPU_W_UNIFORMS_ADDRESS
:
393 DRM_ERROR("relative uniforms address change "
398 return validate_uniform_address_write(validated_shader
,
402 case QPU_W_TLB_COLOR_MS
:
403 case QPU_W_TLB_COLOR_ALL
:
405 /* These only interact with the tile buffer, not main memory,
418 return check_tmu_write(validated_shader
, validation_state
,
422 case QPU_W_TMU_NOSWAP
:
423 case QPU_W_TLB_ALPHA_MASK
:
424 case QPU_W_MUTEX_RELEASE
:
425 /* XXX: I haven't thought about these, so don't support them
428 DRM_ERROR("Unsupported waddr %d\n", waddr
);
432 DRM_ERROR("General VPM DMA unsupported\n");
436 case QPU_W_VPMVCD_SETUP
:
437 /* We allow VPM setup in general, even including VPM DMA
438 * configuration setup, because the (unsafe) DMA can only be
439 * triggered by QPU_W_VPM_ADDR writes.
443 case QPU_W_TLB_STENCIL_SETUP
:
451 track_live_clamps(struct vc4_validated_shader_info
*validated_shader
,
452 struct vc4_shader_validation_state
*validation_state
)
454 uint64_t inst
= validation_state
->shader
[validation_state
->ip
];
455 uint32_t op_add
= QPU_GET_FIELD(inst
, QPU_OP_ADD
);
456 uint32_t waddr_add
= QPU_GET_FIELD(inst
, QPU_WADDR_ADD
);
457 uint32_t waddr_mul
= QPU_GET_FIELD(inst
, QPU_WADDR_MUL
);
458 uint32_t cond_add
= QPU_GET_FIELD(inst
, QPU_COND_ADD
);
459 uint32_t add_a
= QPU_GET_FIELD(inst
, QPU_ADD_A
);
460 uint32_t add_b
= QPU_GET_FIELD(inst
, QPU_ADD_B
);
461 uint32_t raddr_a
= QPU_GET_FIELD(inst
, QPU_RADDR_A
);
462 uint32_t raddr_b
= QPU_GET_FIELD(inst
, QPU_RADDR_B
);
463 uint32_t sig
= QPU_GET_FIELD(inst
, QPU_SIG
);
464 bool ws
= inst
& QPU_WS
;
465 uint32_t lri_add_a
, lri_add
, lri_mul
;
468 /* Check whether OP_ADD's A argumennt comes from a live MAX(x, 0),
469 * before we clear previous live state.
471 lri_add_a
= raddr_add_a_to_live_reg_index(inst
);
472 add_a_is_min_0
= (lri_add_a
!= ~0 &&
473 validation_state
->live_max_clamp_regs
[lri_add_a
]);
475 /* Clear live state for registers written by our instruction. */
476 lri_add
= waddr_to_live_reg_index(waddr_add
, ws
);
477 lri_mul
= waddr_to_live_reg_index(waddr_mul
, !ws
);
479 validation_state
->live_max_clamp_regs
[lri_mul
] = false;
480 validation_state
->live_min_clamp_offsets
[lri_mul
] = ~0;
483 validation_state
->live_max_clamp_regs
[lri_add
] = false;
484 validation_state
->live_min_clamp_offsets
[lri_add
] = ~0;
486 /* Nothing further to do for live tracking, since only ADDs
487 * generate new live clamp registers.
492 /* Now, handle remaining live clamp tracking for the ADD operation. */
494 if (cond_add
!= QPU_COND_ALWAYS
)
497 if (op_add
== QPU_A_MAX
) {
498 /* Track live clamps of a value to a minimum of 0 (in either
501 if (sig
!= QPU_SIG_SMALL_IMM
|| raddr_b
!= 0 ||
502 (add_a
!= QPU_MUX_B
&& add_b
!= QPU_MUX_B
)) {
506 validation_state
->live_max_clamp_regs
[lri_add
] = true;
507 } else if (op_add
== QPU_A_MIN
) {
508 /* Track live clamps of a value clamped to a minimum of 0 and
509 * a maximum of some uniform's offset.
514 if (!(add_b
== QPU_MUX_A
&& raddr_a
== QPU_R_UNIF
) &&
515 !(add_b
== QPU_MUX_B
&& raddr_b
== QPU_R_UNIF
&&
516 sig
!= QPU_SIG_SMALL_IMM
)) {
520 validation_state
->live_min_clamp_offsets
[lri_add
] =
521 validated_shader
->uniforms_size
;
526 check_instruction_writes(struct vc4_validated_shader_info
*validated_shader
,
527 struct vc4_shader_validation_state
*validation_state
)
529 uint64_t inst
= validation_state
->shader
[validation_state
->ip
];
530 uint32_t waddr_add
= QPU_GET_FIELD(inst
, QPU_WADDR_ADD
);
531 uint32_t waddr_mul
= QPU_GET_FIELD(inst
, QPU_WADDR_MUL
);
534 if (is_tmu_write(waddr_add
) && is_tmu_write(waddr_mul
)) {
535 DRM_ERROR("ADD and MUL both set up textures\n");
539 ok
= (check_reg_write(validated_shader
, validation_state
, false) &&
540 check_reg_write(validated_shader
, validation_state
, true));
542 track_live_clamps(validated_shader
, validation_state
);
548 check_branch(uint64_t inst
,
549 struct vc4_validated_shader_info
*validated_shader
,
550 struct vc4_shader_validation_state
*validation_state
,
553 int32_t branch_imm
= QPU_GET_FIELD(inst
, QPU_BRANCH_TARGET
);
554 uint32_t waddr_add
= QPU_GET_FIELD(inst
, QPU_WADDR_ADD
);
555 uint32_t waddr_mul
= QPU_GET_FIELD(inst
, QPU_WADDR_MUL
);
557 if ((int)branch_imm
< 0)
558 validation_state
->needs_uniform_address_for_loop
= true;
560 /* We don't want to have to worry about validation of this, and
561 * there's no need for it.
563 if (waddr_add
!= QPU_W_NOP
|| waddr_mul
!= QPU_W_NOP
) {
564 DRM_ERROR("branch instruction at %d wrote a register.\n",
565 validation_state
->ip
);
573 check_instruction_reads(struct vc4_validated_shader_info
*validated_shader
,
574 struct vc4_shader_validation_state
*validation_state
)
576 uint64_t inst
= validation_state
->shader
[validation_state
->ip
];
577 uint32_t raddr_a
= QPU_GET_FIELD(inst
, QPU_RADDR_A
);
578 uint32_t raddr_b
= QPU_GET_FIELD(inst
, QPU_RADDR_B
);
579 uint32_t sig
= QPU_GET_FIELD(inst
, QPU_SIG
);
581 if (raddr_a
== QPU_R_UNIF
||
582 (raddr_b
== QPU_R_UNIF
&& sig
!= QPU_SIG_SMALL_IMM
)) {
583 /* This can't overflow the uint32_t, because we're reading 8
584 * bytes of instruction to increment by 4 here, so we'd
587 validated_shader
->uniforms_size
+= 4;
589 if (validation_state
->needs_uniform_address_update
) {
590 DRM_ERROR("Uniform read with undefined uniform "
599 /* Make sure that all branches are absolute and point within the shader, and
600 * note their targets for later.
603 vc4_validate_branches(struct vc4_shader_validation_state
*validation_state
)
605 uint32_t max_branch_target
= 0;
606 bool found_shader_end
= false;
608 int shader_end_ip
= 0;
609 int last_branch
= -2;
611 for (ip
= 0; ip
< validation_state
->max_ip
; ip
++) {
612 uint64_t inst
= validation_state
->shader
[ip
];
613 int32_t branch_imm
= QPU_GET_FIELD(inst
, QPU_BRANCH_TARGET
);
614 uint32_t sig
= QPU_GET_FIELD(inst
, QPU_SIG
);
615 uint32_t after_delay_ip
= ip
+ 4;
616 uint32_t branch_target_ip
;
618 if (sig
== QPU_SIG_PROG_END
) {
620 found_shader_end
= true;
624 if (sig
!= QPU_SIG_BRANCH
)
627 if (ip
- last_branch
< 4) {
628 DRM_ERROR("Branch at %d during delay slots\n", ip
);
633 if (inst
& QPU_BRANCH_REG
) {
634 DRM_ERROR("branching from register relative "
639 if (!(inst
& QPU_BRANCH_REL
)) {
640 DRM_ERROR("relative branching required\n");
644 /* The actual branch target is the instruction after the delay
645 * slots, plus whatever byte offset is in the low 32 bits of
646 * the instruction. Make sure we're not branching beyond the
647 * end of the shader object.
649 if (branch_imm
% sizeof(inst
) != 0) {
650 DRM_ERROR("branch target not aligned\n");
654 branch_target_ip
= after_delay_ip
+ (branch_imm
>> 3);
655 if (branch_target_ip
>= validation_state
->max_ip
) {
656 DRM_ERROR("Branch at %d outside of shader (ip %d/%d)\n",
657 ip
, branch_target_ip
,
658 validation_state
->max_ip
);
661 set_bit(branch_target_ip
, validation_state
->branch_targets
);
663 /* Make sure that the non-branching path is also not outside
666 if (after_delay_ip
>= validation_state
->max_ip
) {
667 DRM_ERROR("Branch at %d continues past shader end "
669 ip
, after_delay_ip
, validation_state
->max_ip
);
672 set_bit(after_delay_ip
, validation_state
->branch_targets
);
673 max_branch_target
= max(max_branch_target
, after_delay_ip
);
675 /* There are two delay slots after program end is signaled
676 * that are still executed, then we're finished.
678 if (found_shader_end
&& ip
== shader_end_ip
+ 2)
682 if (max_branch_target
> shader_end_ip
) {
683 DRM_ERROR("Branch landed after QPU_SIG_PROG_END");
690 /* Resets any known state for the shader, used when we may be branched to from
691 * multiple locations in the program (or at shader start).
694 reset_validation_state(struct vc4_shader_validation_state
*validation_state
)
698 for (i
= 0; i
< 8; i
++)
699 validation_state
->tmu_setup
[i
/ 4].p_offset
[i
% 4] = ~0;
701 for (i
= 0; i
< LIVE_REG_COUNT
; i
++) {
702 validation_state
->live_min_clamp_offsets
[i
] = ~0;
703 validation_state
->live_max_clamp_regs
[i
] = false;
704 validation_state
->live_immediates
[i
] = ~0;
709 texturing_in_progress(struct vc4_shader_validation_state
*validation_state
)
711 return (validation_state
->tmu_write_count
[0] != 0 ||
712 validation_state
->tmu_write_count
[1] != 0);
716 vc4_handle_branch_target(struct vc4_shader_validation_state
*validation_state
)
718 uint32_t ip
= validation_state
->ip
;
720 if (!test_bit(ip
, validation_state
->branch_targets
))
723 if (texturing_in_progress(validation_state
)) {
724 DRM_ERROR("Branch target landed during TMU setup\n");
728 /* Reset our live values tracking, since this instruction may have
729 * multiple predecessors.
731 * One could potentially do analysis to determine that, for
732 * example, all predecessors have a live max clamp in the same
733 * register, but we don't bother with that.
735 reset_validation_state(validation_state
);
737 /* Since we've entered a basic block from potentially multiple
738 * predecessors, we need the uniforms address to be updated before any
739 * unforms are read. We require that after any branch point, the next
740 * uniform to be loaded is a uniform address offset. That uniform's
741 * offset will be marked by the uniform address register write
742 * validation, or a one-off the end-of-program check.
744 validation_state
->needs_uniform_address_update
= true;
749 struct vc4_validated_shader_info
*
750 vc4_validate_shader(struct drm_gem_cma_object
*shader_obj
)
752 bool found_shader_end
= false;
753 int shader_end_ip
= 0;
755 struct vc4_validated_shader_info
*validated_shader
= NULL
;
756 struct vc4_shader_validation_state validation_state
;
758 memset(&validation_state
, 0, sizeof(validation_state
));
759 validation_state
.shader
= shader_obj
->vaddr
;
760 validation_state
.max_ip
= shader_obj
->base
.size
/ sizeof(uint64_t);
762 reset_validation_state(&validation_state
);
764 validation_state
.branch_targets
=
765 kcalloc(BITS_TO_LONGS(validation_state
.max_ip
),
766 sizeof(unsigned long), GFP_KERNEL
);
767 if (!validation_state
.branch_targets
)
770 validated_shader
= kcalloc(1, sizeof(*validated_shader
), GFP_KERNEL
);
771 if (!validated_shader
)
774 if (!vc4_validate_branches(&validation_state
))
777 for (ip
= 0; ip
< validation_state
.max_ip
; ip
++) {
778 uint64_t inst
= validation_state
.shader
[ip
];
779 uint32_t sig
= QPU_GET_FIELD(inst
, QPU_SIG
);
781 validation_state
.ip
= ip
;
783 if (!vc4_handle_branch_target(&validation_state
))
788 case QPU_SIG_WAIT_FOR_SCOREBOARD
:
789 case QPU_SIG_SCOREBOARD_UNLOCK
:
790 case QPU_SIG_COLOR_LOAD
:
791 case QPU_SIG_LOAD_TMU0
:
792 case QPU_SIG_LOAD_TMU1
:
793 case QPU_SIG_PROG_END
:
794 case QPU_SIG_SMALL_IMM
:
795 if (!check_instruction_writes(validated_shader
,
796 &validation_state
)) {
797 DRM_ERROR("Bad write at ip %d\n", ip
);
801 if (!check_instruction_reads(validated_shader
,
805 if (sig
== QPU_SIG_PROG_END
) {
806 found_shader_end
= true;
812 case QPU_SIG_LOAD_IMM
:
813 if (!check_instruction_writes(validated_shader
,
814 &validation_state
)) {
815 DRM_ERROR("Bad LOAD_IMM write at ip %d\n", ip
);
821 if (!check_branch(inst
, validated_shader
,
822 &validation_state
, ip
))
826 DRM_ERROR("Unsupported QPU signal %d at "
827 "instruction %d\n", sig
, ip
);
831 /* There are two delay slots after program end is signaled
832 * that are still executed, then we're finished.
834 if (found_shader_end
&& ip
== shader_end_ip
+ 2)
838 if (ip
== validation_state
.max_ip
) {
839 DRM_ERROR("shader failed to terminate before "
840 "shader BO end at %zd\n",
841 shader_obj
->base
.size
);
845 /* If we did a backwards branch and we haven't emitted a uniforms
846 * reset since then, we still need the uniforms stream to have the
847 * uniforms address available so that the backwards branch can do its
850 * We could potentially prove that the backwards branch doesn't
851 * contain any uses of uniforms until program exit, but that doesn't
852 * seem to be worth the trouble.
854 if (validation_state
.needs_uniform_address_for_loop
) {
855 if (!require_uniform_address_uniform(validated_shader
))
857 validated_shader
->uniforms_size
+= 4;
860 /* Again, no chance of integer overflow here because the worst case
861 * scenario is 8 bytes of uniforms plus handles per 8-byte
864 validated_shader
->uniforms_src_size
=
865 (validated_shader
->uniforms_size
+
866 4 * validated_shader
->num_texture_samples
);
868 kfree(validation_state
.branch_targets
);
870 return validated_shader
;
873 kfree(validation_state
.branch_targets
);
874 if (validated_shader
) {
875 kfree(validated_shader
->texture_samples
);
876 kfree(validated_shader
);