2 * Copyright 2015 Advanced Micro Devices, Inc.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Marek Olšák <maraeo@gmail.com>
28 #include "si_shader.h"
30 #include "sid_tables.h"
31 #include "radeon/radeon_elf_util.h"
32 #include "ddebug/dd_util.h"
33 #include "util/u_memory.h"
35 DEBUG_GET_ONCE_OPTION(replace_shaders
, "RADEON_REPLACE_SHADERS", NULL
)
37 static void si_dump_shader(struct si_screen
*sscreen
,
38 struct si_shader_ctx_state
*state
, FILE *f
)
40 if (!state
->cso
|| !state
->current
)
43 si_dump_shader_key(state
->cso
->type
, &state
->current
->key
, f
);
44 si_shader_dump(sscreen
, state
->current
, NULL
,
45 state
->cso
->info
.processor
, f
);
49 * Shader compiles can be overridden with arbitrary ELF objects by setting
50 * the environment variable RADEON_REPLACE_SHADERS=num1:filename1[;num2:filename2]
52 bool si_replace_shader(unsigned num
, struct radeon_shader_binary
*binary
)
54 const char *p
= debug_get_option_replace_shaders();
55 const char *semicolon
;
60 bool replaced
= false;
68 i
= strtoul(p
, &endp
, 0);
72 fprintf(stderr
, "RADEON_REPLACE_SHADERS formatted badly.\n");
88 semicolon
= strchr(p
, ';');
90 p
= copy
= strndup(p
, semicolon
- p
);
92 fprintf(stderr
, "out of memory\n");
97 fprintf(stderr
, "radeonsi: replace shader %u by %s\n", num
, p
);
101 perror("radeonsi: failed to open file");
105 if (fseek(f
, 0, SEEK_END
) != 0)
112 if (fseek(f
, 0, SEEK_SET
) != 0)
115 buf
= MALLOC(filesize
);
117 fprintf(stderr
, "out of memory\n");
121 nread
= fread(buf
, 1, filesize
, f
);
122 if (nread
!= filesize
)
125 radeon_elf_read(buf
, filesize
, binary
);
136 perror("radeonsi: reading shader");
140 /* Parsed IBs are difficult to read without colors. Use "less -R file" to
141 * read them, or use "aha -b -f file" to convert them to html.
143 #define COLOR_RESET "\033[0m"
144 #define COLOR_RED "\033[31m"
145 #define COLOR_GREEN "\033[1;32m"
146 #define COLOR_YELLOW "\033[1;33m"
147 #define COLOR_CYAN "\033[1;36m"
151 static void print_spaces(FILE *f
, unsigned num
)
153 fprintf(f
, "%*s", num
, "");
156 static void print_value(FILE *file
, uint32_t value
, int bits
)
158 /* Guess if it's int or float */
159 if (value
<= (1 << 15)) {
161 fprintf(file
, "%u\n", value
);
163 fprintf(file
, "%u (0x%0*x)\n", value
, bits
/ 4, value
);
165 float f
= uif(value
);
167 if (fabs(f
) < 100000 && f
*10 == floor(f
*10))
168 fprintf(file
, "%.1ff (0x%0*x)\n", f
, bits
/ 4, value
);
170 /* Don't print more leading zeros than there are bits. */
171 fprintf(file
, "0x%0*x\n", bits
/ 4, value
);
175 static void print_named_value(FILE *file
, const char *name
, uint32_t value
,
178 print_spaces(file
, INDENT_PKT
);
179 fprintf(file
, COLOR_YELLOW
"%s" COLOR_RESET
" <- ", name
);
180 print_value(file
, value
, bits
);
183 static void si_dump_reg(FILE *file
, unsigned offset
, uint32_t value
,
188 for (r
= 0; r
< ARRAY_SIZE(reg_table
); r
++) {
189 const struct si_reg
*reg
= ®_table
[r
];
191 if (reg
->offset
== offset
) {
192 bool first_field
= true;
194 print_spaces(file
, INDENT_PKT
);
195 fprintf(file
, COLOR_YELLOW
"%s" COLOR_RESET
" <- ",
198 if (!reg
->num_fields
) {
199 print_value(file
, value
, 32);
203 for (f
= 0; f
< reg
->num_fields
; f
++) {
204 const struct si_field
*field
= ®
->fields
[f
];
205 uint32_t val
= (value
& field
->mask
) >>
206 (ffs(field
->mask
) - 1);
208 if (!(field
->mask
& field_mask
))
211 /* Indent the field. */
214 INDENT_PKT
+ strlen(reg
->name
) + 4);
216 /* Print the field. */
217 fprintf(file
, "%s = ", field
->name
);
219 if (val
< field
->num_values
&& field
->values
[val
])
220 fprintf(file
, "%s\n", field
->values
[val
]);
222 print_value(file
, val
,
223 util_bitcount(field
->mask
));
231 fprintf(file
, COLOR_YELLOW
"0x%05x" COLOR_RESET
" = 0x%08x", offset
, value
);
234 static void si_parse_set_reg_packet(FILE *f
, uint32_t *ib
, unsigned count
,
237 unsigned reg
= (ib
[1] << 2) + reg_offset
;
240 for (i
= 0; i
< count
; i
++)
241 si_dump_reg(f
, reg
+ i
*4, ib
[2+i
], ~0);
244 static uint32_t *si_parse_packet3(FILE *f
, uint32_t *ib
, int *num_dw
,
247 unsigned count
= PKT_COUNT_G(ib
[0]);
248 unsigned op
= PKT3_IT_OPCODE_G(ib
[0]);
249 const char *predicate
= PKT3_PREDICATE(ib
[0]) ? "(predicate)" : "";
252 /* Print the name first. */
253 for (i
= 0; i
< ARRAY_SIZE(packet3_table
); i
++)
254 if (packet3_table
[i
].op
== op
)
257 if (i
< ARRAY_SIZE(packet3_table
))
258 if (op
== PKT3_SET_CONTEXT_REG
||
259 op
== PKT3_SET_CONFIG_REG
||
260 op
== PKT3_SET_UCONFIG_REG
||
261 op
== PKT3_SET_SH_REG
)
262 fprintf(f
, COLOR_CYAN
"%s%s" COLOR_CYAN
":\n",
263 packet3_table
[i
].name
, predicate
);
265 fprintf(f
, COLOR_GREEN
"%s%s" COLOR_RESET
":\n",
266 packet3_table
[i
].name
, predicate
);
268 fprintf(f
, COLOR_RED
"PKT3_UNKNOWN 0x%x%s" COLOR_RESET
":\n",
271 /* Print the contents. */
273 case PKT3_SET_CONTEXT_REG
:
274 si_parse_set_reg_packet(f
, ib
, count
, SI_CONTEXT_REG_OFFSET
);
276 case PKT3_SET_CONFIG_REG
:
277 si_parse_set_reg_packet(f
, ib
, count
, SI_CONFIG_REG_OFFSET
);
279 case PKT3_SET_UCONFIG_REG
:
280 si_parse_set_reg_packet(f
, ib
, count
, CIK_UCONFIG_REG_OFFSET
);
282 case PKT3_SET_SH_REG
:
283 si_parse_set_reg_packet(f
, ib
, count
, SI_SH_REG_OFFSET
);
285 case PKT3_DRAW_PREAMBLE
:
286 si_dump_reg(f
, R_030908_VGT_PRIMITIVE_TYPE
, ib
[1], ~0);
287 si_dump_reg(f
, R_028AA8_IA_MULTI_VGT_PARAM
, ib
[2], ~0);
288 si_dump_reg(f
, R_028B58_VGT_LS_HS_CONFIG
, ib
[3], ~0);
290 case PKT3_ACQUIRE_MEM
:
291 si_dump_reg(f
, R_0301F0_CP_COHER_CNTL
, ib
[1], ~0);
292 si_dump_reg(f
, R_0301F4_CP_COHER_SIZE
, ib
[2], ~0);
293 si_dump_reg(f
, R_030230_CP_COHER_SIZE_HI
, ib
[3], ~0);
294 si_dump_reg(f
, R_0301F8_CP_COHER_BASE
, ib
[4], ~0);
295 si_dump_reg(f
, R_0301E4_CP_COHER_BASE_HI
, ib
[5], ~0);
296 print_named_value(f
, "POLL_INTERVAL", ib
[6], 16);
298 case PKT3_SURFACE_SYNC
:
299 si_dump_reg(f
, R_0085F0_CP_COHER_CNTL
, ib
[1], ~0);
300 si_dump_reg(f
, R_0085F4_CP_COHER_SIZE
, ib
[2], ~0);
301 si_dump_reg(f
, R_0085F8_CP_COHER_BASE
, ib
[3], ~0);
302 print_named_value(f
, "POLL_INTERVAL", ib
[4], 16);
304 case PKT3_EVENT_WRITE
:
305 si_dump_reg(f
, R_028A90_VGT_EVENT_INITIATOR
, ib
[1],
306 S_028A90_EVENT_TYPE(~0));
307 print_named_value(f
, "EVENT_INDEX", (ib
[1] >> 8) & 0xf, 4);
308 print_named_value(f
, "INV_L2", (ib
[1] >> 20) & 0x1, 1);
310 print_named_value(f
, "ADDRESS_LO", ib
[2], 32);
311 print_named_value(f
, "ADDRESS_HI", ib
[3], 16);
314 case PKT3_DRAW_INDEX_AUTO
:
315 si_dump_reg(f
, R_030930_VGT_NUM_INDICES
, ib
[1], ~0);
316 si_dump_reg(f
, R_0287F0_VGT_DRAW_INITIATOR
, ib
[2], ~0);
318 case PKT3_DRAW_INDEX_2
:
319 si_dump_reg(f
, R_028A78_VGT_DMA_MAX_SIZE
, ib
[1], ~0);
320 si_dump_reg(f
, R_0287E8_VGT_DMA_BASE
, ib
[2], ~0);
321 si_dump_reg(f
, R_0287E4_VGT_DMA_BASE_HI
, ib
[3], ~0);
322 si_dump_reg(f
, R_030930_VGT_NUM_INDICES
, ib
[4], ~0);
323 si_dump_reg(f
, R_0287F0_VGT_DRAW_INITIATOR
, ib
[5], ~0);
325 case PKT3_INDEX_TYPE
:
326 si_dump_reg(f
, R_028A7C_VGT_DMA_INDEX_TYPE
, ib
[1], ~0);
328 case PKT3_NUM_INSTANCES
:
329 si_dump_reg(f
, R_030934_VGT_NUM_INSTANCES
, ib
[1], ~0);
331 case PKT3_WRITE_DATA
:
332 si_dump_reg(f
, R_370_CONTROL
, ib
[1], ~0);
333 si_dump_reg(f
, R_371_DST_ADDR_LO
, ib
[2], ~0);
334 si_dump_reg(f
, R_372_DST_ADDR_HI
, ib
[3], ~0);
335 for (i
= 2; i
< count
; i
++) {
336 print_spaces(f
, INDENT_PKT
);
337 fprintf(f
, "0x%08x\n", ib
[2+i
]);
341 si_dump_reg(f
, R_410_CP_DMA_WORD0
, ib
[1], ~0);
342 si_dump_reg(f
, R_411_CP_DMA_WORD1
, ib
[2], ~0);
343 si_dump_reg(f
, R_412_CP_DMA_WORD2
, ib
[3], ~0);
344 si_dump_reg(f
, R_413_CP_DMA_WORD3
, ib
[4], ~0);
345 si_dump_reg(f
, R_414_COMMAND
, ib
[5], ~0);
348 si_dump_reg(f
, R_500_DMA_DATA_WORD0
, ib
[1], ~0);
349 si_dump_reg(f
, R_501_SRC_ADDR_LO
, ib
[2], ~0);
350 si_dump_reg(f
, R_502_SRC_ADDR_HI
, ib
[3], ~0);
351 si_dump_reg(f
, R_503_DST_ADDR_LO
, ib
[4], ~0);
352 si_dump_reg(f
, R_504_DST_ADDR_HI
, ib
[5], ~0);
353 si_dump_reg(f
, R_414_COMMAND
, ib
[6], ~0);
356 if (ib
[0] == 0xffff1000) {
357 count
= -1; /* One dword NOP. */
359 } else if (count
== 0 && SI_IS_TRACE_POINT(ib
[1])) {
360 unsigned packet_id
= SI_GET_TRACE_POINT_ID(ib
[1]);
362 print_spaces(f
, INDENT_PKT
);
363 fprintf(f
, COLOR_RED
"Trace point ID: %u\n", packet_id
);
366 break; /* tracing was disabled */
368 print_spaces(f
, INDENT_PKT
);
369 if (packet_id
< trace_id
)
371 "This trace point was reached by the CP."
373 else if (packet_id
== trace_id
)
375 "!!!!! This is the last trace point that "
376 "was reached by the CP !!!!!"
378 else if (packet_id
+1 == trace_id
)
380 "!!!!! This is the first trace point that "
381 "was NOT been reached by the CP !!!!!"
385 "!!!!! This trace point was NOT reached "
390 /* fall through, print all dwords */
392 for (i
= 0; i
< count
+1; i
++) {
393 print_spaces(f
, INDENT_PKT
);
394 fprintf(f
, "0x%08x\n", ib
[1+i
]);
399 *num_dw
-= count
+ 2;
404 * Parse and print an IB into a file.
408 * \param num_dw size of the IB
409 * \param chip_class chip class
410 * \param trace_id the last trace ID that is known to have been reached
411 * and executed by the CP, typically read from a buffer
413 static void si_parse_ib(FILE *f
, uint32_t *ib
, int num_dw
, int trace_id
,
416 fprintf(f
, "------------------ %s begin ------------------\n", name
);
419 unsigned type
= PKT_TYPE_G(ib
[0]);
423 ib
= si_parse_packet3(f
, ib
, &num_dw
, trace_id
);
427 if (ib
[0] == 0x80000000) {
428 fprintf(f
, COLOR_GREEN
"NOP (type 2)" COLOR_RESET
"\n");
434 fprintf(f
, "Unknown packet type %i\n", type
);
439 fprintf(f
, "------------------- %s end -------------------\n", name
);
441 printf("Packet ends after the end of IB.\n");
447 static void si_dump_mmapped_reg(struct si_context
*sctx
, FILE *f
,
450 struct radeon_winsys
*ws
= sctx
->b
.ws
;
453 if (ws
->read_registers(ws
, offset
, 1, &value
))
454 si_dump_reg(f
, offset
, value
, ~0);
457 static void si_dump_debug_registers(struct si_context
*sctx
, FILE *f
)
459 if (sctx
->screen
->b
.info
.drm_major
== 2 &&
460 sctx
->screen
->b
.info
.drm_minor
< 42)
461 return; /* no radeon support */
463 fprintf(f
, "Memory-mapped registers:\n");
464 si_dump_mmapped_reg(sctx
, f
, R_008010_GRBM_STATUS
);
466 /* No other registers can be read on DRM < 3.1.0. */
467 if (sctx
->screen
->b
.info
.drm_major
< 3 ||
468 sctx
->screen
->b
.info
.drm_minor
< 1) {
473 si_dump_mmapped_reg(sctx
, f
, R_008008_GRBM_STATUS2
);
474 si_dump_mmapped_reg(sctx
, f
, R_008014_GRBM_STATUS_SE0
);
475 si_dump_mmapped_reg(sctx
, f
, R_008018_GRBM_STATUS_SE1
);
476 si_dump_mmapped_reg(sctx
, f
, R_008038_GRBM_STATUS_SE2
);
477 si_dump_mmapped_reg(sctx
, f
, R_00803C_GRBM_STATUS_SE3
);
478 si_dump_mmapped_reg(sctx
, f
, R_00D034_SDMA0_STATUS_REG
);
479 si_dump_mmapped_reg(sctx
, f
, R_00D834_SDMA1_STATUS_REG
);
480 si_dump_mmapped_reg(sctx
, f
, R_000E50_SRBM_STATUS
);
481 si_dump_mmapped_reg(sctx
, f
, R_000E4C_SRBM_STATUS2
);
482 si_dump_mmapped_reg(sctx
, f
, R_000E54_SRBM_STATUS3
);
483 si_dump_mmapped_reg(sctx
, f
, R_008680_CP_STAT
);
484 si_dump_mmapped_reg(sctx
, f
, R_008674_CP_STALLED_STAT1
);
485 si_dump_mmapped_reg(sctx
, f
, R_008678_CP_STALLED_STAT2
);
486 si_dump_mmapped_reg(sctx
, f
, R_008670_CP_STALLED_STAT3
);
487 si_dump_mmapped_reg(sctx
, f
, R_008210_CP_CPC_STATUS
);
488 si_dump_mmapped_reg(sctx
, f
, R_008214_CP_CPC_BUSY_STAT
);
489 si_dump_mmapped_reg(sctx
, f
, R_008218_CP_CPC_STALLED_STAT1
);
490 si_dump_mmapped_reg(sctx
, f
, R_00821C_CP_CPF_STATUS
);
491 si_dump_mmapped_reg(sctx
, f
, R_008220_CP_CPF_BUSY_STAT
);
492 si_dump_mmapped_reg(sctx
, f
, R_008224_CP_CPF_STALLED_STAT1
);
496 static void si_dump_last_ib(struct si_context
*sctx
, FILE *f
)
498 int last_trace_id
= -1;
503 if (sctx
->last_trace_buf
) {
504 /* We are expecting that the ddebug pipe has already
505 * waited for the context, so this buffer should be idle.
506 * If the GPU is hung, there is no point in waiting for it.
508 uint32_t *map
= sctx
->b
.ws
->buffer_map(sctx
->last_trace_buf
->buf
,
510 PIPE_TRANSFER_UNSYNCHRONIZED
|
513 last_trace_id
= *map
;
516 if (sctx
->init_config
)
517 si_parse_ib(f
, sctx
->init_config
->pm4
, sctx
->init_config
->ndw
,
518 -1, "IB2: Init config");
520 if (sctx
->init_config_gs_rings
)
521 si_parse_ib(f
, sctx
->init_config_gs_rings
->pm4
,
522 sctx
->init_config_gs_rings
->ndw
,
523 -1, "IB2: Init GS rings");
525 si_parse_ib(f
, sctx
->last_ib
, sctx
->last_ib_dw_size
,
526 last_trace_id
, "IB");
527 free(sctx
->last_ib
); /* dump only once */
528 sctx
->last_ib
= NULL
;
529 r600_resource_reference(&sctx
->last_trace_buf
, NULL
);
532 static const char *priority_to_string(enum radeon_bo_priority priority
)
534 #define ITEM(x) [RADEON_PRIO_##x] = #x
535 static const char *table
[64] = {
538 ITEM(SO_FILLED_SIZE
),
550 ITEM(INTERNAL_SHADER
),
554 ITEM(SAMPLER_BUFFER
),
556 ITEM(SHADER_RW_BUFFER
),
557 ITEM(RINGS_STREAMOUT
),
558 ITEM(SCRATCH_BUFFER
),
559 ITEM(COMPUTE_GLOBAL
),
560 ITEM(SAMPLER_TEXTURE
),
561 ITEM(SHADER_RW_IMAGE
),
562 ITEM(SAMPLER_TEXTURE_MSAA
),
565 ITEM(COLOR_BUFFER_MSAA
),
566 ITEM(DEPTH_BUFFER_MSAA
),
573 assert(priority
< ARRAY_SIZE(table
));
574 return table
[priority
];
577 static int bo_list_compare_va(const struct radeon_bo_list_item
*a
,
578 const struct radeon_bo_list_item
*b
)
580 return a
->vm_address
< b
->vm_address
? -1 :
581 a
->vm_address
> b
->vm_address
? 1 : 0;
584 static void si_dump_last_bo_list(struct si_context
*sctx
, FILE *f
)
588 if (!sctx
->last_bo_list
)
591 /* Sort the list according to VM adddresses first. */
592 qsort(sctx
->last_bo_list
, sctx
->last_bo_count
,
593 sizeof(sctx
->last_bo_list
[0]), (void*)bo_list_compare_va
);
595 fprintf(f
, "Buffer list (in units of pages = 4kB):\n"
596 COLOR_YELLOW
" Size VM start page "
597 "VM end page Usage" COLOR_RESET
"\n");
599 for (i
= 0; i
< sctx
->last_bo_count
; i
++) {
600 /* Note: Buffer sizes are expected to be aligned to 4k by the winsys. */
601 const unsigned page_size
= 4096;
602 uint64_t va
= sctx
->last_bo_list
[i
].vm_address
;
603 uint64_t size
= sctx
->last_bo_list
[i
].buf
->size
;
606 /* If there's unused virtual memory between 2 buffers, print it. */
608 uint64_t previous_va_end
= sctx
->last_bo_list
[i
-1].vm_address
+
609 sctx
->last_bo_list
[i
-1].buf
->size
;
611 if (va
> previous_va_end
) {
612 fprintf(f
, " %10"PRIu64
" -- hole --\n",
613 (va
- previous_va_end
) / page_size
);
617 /* Print the buffer. */
618 fprintf(f
, " %10"PRIu64
" 0x%013"PRIx64
" 0x%013"PRIx64
" ",
619 size
/ page_size
, va
/ page_size
, (va
+ size
) / page_size
);
621 /* Print the usage. */
622 for (j
= 0; j
< 64; j
++) {
623 if (!(sctx
->last_bo_list
[i
].priority_usage
& (1llu << j
)))
626 fprintf(f
, "%s%s", !hit
? "" : ", ", priority_to_string(j
));
631 fprintf(f
, "\nNote: The holes represent memory not used by the IB.\n"
632 " Other buffers can still be allocated there.\n\n");
634 for (i
= 0; i
< sctx
->last_bo_count
; i
++)
635 pb_reference(&sctx
->last_bo_list
[i
].buf
, NULL
);
636 free(sctx
->last_bo_list
);
637 sctx
->last_bo_list
= NULL
;
640 static void si_dump_framebuffer(struct si_context
*sctx
, FILE *f
)
642 struct pipe_framebuffer_state
*state
= &sctx
->framebuffer
.state
;
643 struct r600_texture
*rtex
;
646 for (i
= 0; i
< state
->nr_cbufs
; i
++) {
647 if (!state
->cbufs
[i
])
650 rtex
= (struct r600_texture
*)state
->cbufs
[i
]->texture
;
651 fprintf(f
, COLOR_YELLOW
"Color buffer %i:" COLOR_RESET
"\n", i
);
652 r600_print_texture_info(rtex
, f
);
657 rtex
= (struct r600_texture
*)state
->zsbuf
->texture
;
658 fprintf(f
, COLOR_YELLOW
"Depth-stencil buffer:" COLOR_RESET
"\n");
659 r600_print_texture_info(rtex
, f
);
664 static void si_dump_debug_state(struct pipe_context
*ctx
, FILE *f
,
667 struct si_context
*sctx
= (struct si_context
*)ctx
;
669 if (flags
& PIPE_DEBUG_DEVICE_IS_HUNG
)
670 si_dump_debug_registers(sctx
, f
);
672 si_dump_framebuffer(sctx
, f
);
673 si_dump_shader(sctx
->screen
, &sctx
->vs_shader
, f
);
674 si_dump_shader(sctx
->screen
, &sctx
->tcs_shader
, f
);
675 si_dump_shader(sctx
->screen
, &sctx
->tes_shader
, f
);
676 si_dump_shader(sctx
->screen
, &sctx
->gs_shader
, f
);
677 si_dump_shader(sctx
->screen
, &sctx
->ps_shader
, f
);
679 si_dump_last_bo_list(sctx
, f
);
680 si_dump_last_ib(sctx
, f
);
682 fprintf(f
, "Done.\n");
685 static bool si_vm_fault_occured(struct si_context
*sctx
, uint32_t *out_addr
)
690 uint64_t timestamp
= 0;
693 FILE *p
= popen("dmesg", "r");
697 while (fgets(line
, sizeof(line
), p
)) {
700 /* Get the timestamp. */
701 if (sscanf(line
, "[%u.%u]", &sec
, &usec
) != 2) {
705 timestamp
= sec
* 1000000llu + usec
;
707 /* If just updating the timestamp. */
711 /* Process messages only if the timestamp is newer. */
712 if (timestamp
<= sctx
->dmesg_timestamp
)
715 /* Only process the first VM fault. */
719 /* Remove trailing \n */
721 if (len
&& line
[len
-1] == '\n')
724 /* Get the message part. */
725 msg
= strchr(line
, ']');
734 if (strstr(msg
, "GPU fault detected:"))
738 msg
= strstr(msg
, "VM_CONTEXT1_PROTECTION_FAULT_ADDR");
740 msg
= strstr(msg
, "0x");
743 if (sscanf(msg
, "%X", out_addr
) == 1)
755 if (timestamp
> sctx
->dmesg_timestamp
)
756 sctx
->dmesg_timestamp
= timestamp
;
760 void si_check_vm_faults(struct si_context
*sctx
)
762 struct pipe_screen
*screen
= sctx
->b
.b
.screen
;
766 /* Use conservative timeout 800ms, after which we won't wait any
767 * longer and assume the GPU is hung.
769 sctx
->b
.ws
->fence_wait(sctx
->b
.ws
, sctx
->last_gfx_fence
, 800*1000*1000);
771 if (!si_vm_fault_occured(sctx
, &addr
))
774 f
= dd_get_debug_file(false);
778 fprintf(f
, "VM fault report.\n\n");
779 fprintf(f
, "Driver vendor: %s\n", screen
->get_vendor(screen
));
780 fprintf(f
, "Device vendor: %s\n", screen
->get_device_vendor(screen
));
781 fprintf(f
, "Device name: %s\n\n", screen
->get_name(screen
));
782 fprintf(f
, "Failing VM page: 0x%08x\n\n", addr
);
784 si_dump_debug_state(&sctx
->b
.b
, f
, 0);
787 fprintf(stderr
, "Detected a VM fault, exiting...\n");
791 void si_init_debug_functions(struct si_context
*sctx
)
793 sctx
->b
.b
.dump_debug_state
= si_dump_debug_state
;
795 /* Set the initial dmesg timestamp for this context, so that
796 * only new messages will be checked for VM faults.
798 if (sctx
->screen
->b
.debug_flags
& DBG_CHECK_VM
)
799 si_vm_fault_occured(sctx
, NULL
);