deadlock_tests.c revision 41687f09
1/*
2 * Copyright 2017 Advanced Micro Devices, Inc.
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 shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22*/
23
24#include <stdio.h>
25#include <stdlib.h>
26#include <unistd.h>
27#if HAVE_ALLOCA_H
28# include <alloca.h>
29#endif
30
31#include "CUnit/Basic.h"
32
33#include "amdgpu_test.h"
34#include "amdgpu_drm.h"
35#include "amdgpu_internal.h"
36
37#include <pthread.h>
38
39
40/*
41 * This defines the delay in MS after which memory location designated for
42 * compression against reference value is written to, unblocking command
43 * processor
44 */
45#define WRITE_MEM_ADDRESS_DELAY_MS 100
46
47#define	PACKET_TYPE3	3
48
49#define PACKET3(op, n)	((PACKET_TYPE3 << 30) |				\
50			 (((op) & 0xFF) << 8) |				\
51			 ((n) & 0x3FFF) << 16)
52
53#define	PACKET3_WAIT_REG_MEM				0x3C
54#define		WAIT_REG_MEM_FUNCTION(x)                ((x) << 0)
55		/* 0 - always
56		 * 1 - <
57		 * 2 - <=
58		 * 3 - ==
59		 * 4 - !=
60		 * 5 - >=
61		 * 6 - >
62		 */
63#define		WAIT_REG_MEM_MEM_SPACE(x)               ((x) << 4)
64		/* 0 - reg
65		 * 1 - mem
66		 */
67#define		WAIT_REG_MEM_OPERATION(x)               ((x) << 6)
68		/* 0 - wait_reg_mem
69		 * 1 - wr_wait_wr_reg
70		 */
71#define		WAIT_REG_MEM_ENGINE(x)                  ((x) << 8)
72		/* 0 - me
73		 * 1 - pfp
74		 */
75
76#define	PACKET3_WRITE_DATA				0x37
77#define		WRITE_DATA_DST_SEL(x)                   ((x) << 8)
78		/* 0 - register
79		 * 1 - memory (sync - via GRBM)
80		 * 2 - gl2
81		 * 3 - gds
82		 * 4 - reserved
83		 * 5 - memory (async - direct)
84		 */
85#define		WR_ONE_ADDR                             (1 << 16)
86#define		WR_CONFIRM                              (1 << 20)
87#define		WRITE_DATA_CACHE_POLICY(x)              ((x) << 25)
88		/* 0 - LRU
89		 * 1 - Stream
90		 */
91#define		WRITE_DATA_ENGINE_SEL(x)                ((x) << 30)
92		/* 0 - me
93		 * 1 - pfp
94		 * 2 - ce
95		 */
96
97#define mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR                                      0x54f
98
99#define SDMA_PKT_HEADER_OP(x)	(x & 0xff)
100#define SDMA_OP_POLL_REGMEM  8
101
102static  amdgpu_device_handle device_handle;
103static  uint32_t  major_version;
104static  uint32_t  minor_version;
105
106static pthread_t stress_thread;
107static uint32_t *ptr;
108
109int use_uc_mtype = 0;
110
111static void amdgpu_deadlock_helper(unsigned ip_type);
112static void amdgpu_deadlock_gfx(void);
113static void amdgpu_deadlock_compute(void);
114static void amdgpu_illegal_reg_access();
115static void amdgpu_illegal_mem_access();
116static void amdgpu_deadlock_sdma(void);
117static void amdgpu_dispatch_hang_gfx(void);
118static void amdgpu_dispatch_hang_compute(void);
119static void amdgpu_dispatch_hang_slow_gfx(void);
120static void amdgpu_dispatch_hang_slow_compute(void);
121static void amdgpu_draw_hang_gfx(void);
122static void amdgpu_draw_hang_slow_gfx(void);
123
124CU_BOOL suite_deadlock_tests_enable(void)
125{
126	CU_BOOL enable = CU_TRUE;
127	uint32_t asic_id;
128
129	if (amdgpu_device_initialize(drm_amdgpu[0], &major_version,
130					     &minor_version, &device_handle))
131		return CU_FALSE;
132
133	/*
134	 * Only enable for ASICs supporting GPU reset and for which it's enabled
135	 * by default (currently GFX8/9 dGPUS)
136	 */
137	if (device_handle->info.family_id != AMDGPU_FAMILY_VI &&
138	    device_handle->info.family_id != AMDGPU_FAMILY_AI &&
139	    device_handle->info.family_id != AMDGPU_FAMILY_CI) {
140		printf("\n\nGPU reset is not enabled for the ASIC, deadlock suite disabled\n");
141		enable = CU_FALSE;
142	}
143
144	asic_id = device_handle->info.asic_id;
145	if (asic_is_arcturus(asic_id)) {
146		if (amdgpu_set_test_active("Deadlock Tests",
147					"gfx ring block test (set amdgpu.lockup_timeout=50)",
148					CU_FALSE))
149			fprintf(stderr, "test deactivation failed - %s\n",
150				CU_get_error_msg());
151	}
152
153	if (device_handle->info.family_id >= AMDGPU_FAMILY_AI)
154		use_uc_mtype = 1;
155
156	if (amdgpu_device_deinitialize(device_handle))
157		return CU_FALSE;
158
159	return enable;
160}
161
162int suite_deadlock_tests_init(void)
163{
164	int r;
165
166	r = amdgpu_device_initialize(drm_amdgpu[0], &major_version,
167				   &minor_version, &device_handle);
168
169	if (r) {
170		if ((r == -EACCES) && (errno == EACCES))
171			printf("\n\nError:%s. "
172				"Hint:Try to run this test program as root.",
173				strerror(errno));
174		return CUE_SINIT_FAILED;
175	}
176
177	return CUE_SUCCESS;
178}
179
180int suite_deadlock_tests_clean(void)
181{
182	int r = amdgpu_device_deinitialize(device_handle);
183
184	if (r == 0)
185		return CUE_SUCCESS;
186	else
187		return CUE_SCLEAN_FAILED;
188}
189
190
191CU_TestInfo deadlock_tests[] = {
192	{ "gfx ring block test (set amdgpu.lockup_timeout=50)", amdgpu_deadlock_gfx },
193	{ "compute ring block test (set amdgpu.lockup_timeout=50)", amdgpu_deadlock_compute },
194	{ "sdma ring block test (set amdgpu.lockup_timeout=50)", amdgpu_deadlock_sdma },
195	{ "illegal reg access test", amdgpu_illegal_reg_access },
196	{ "illegal mem access test (set amdgpu.vm_fault_stop=2)", amdgpu_illegal_mem_access },
197	{ "gfx ring bad dispatch test (set amdgpu.lockup_timeout=50)", amdgpu_dispatch_hang_gfx },
198	{ "compute ring bad dispatch test (set amdgpu.lockup_timeout=50,50)", amdgpu_dispatch_hang_compute },
199	{ "gfx ring bad slow dispatch test (set amdgpu.lockup_timeout=50)", amdgpu_dispatch_hang_slow_gfx },
200	{ "compute ring bad slow dispatch test (set amdgpu.lockup_timeout=50,50)", amdgpu_dispatch_hang_slow_compute },
201	{ "gfx ring bad draw test (set amdgpu.lockup_timeout=50)", amdgpu_draw_hang_gfx },
202	{ "gfx ring slow bad draw test (set amdgpu.lockup_timeout=50)", amdgpu_draw_hang_slow_gfx },
203	CU_TEST_INFO_NULL,
204};
205
206static void *write_mem_address(void *data)
207{
208	int i;
209
210	/* useconds_t range is [0, 1,000,000] so use loop for waits > 1s */
211	for (i = 0; i < WRITE_MEM_ADDRESS_DELAY_MS; i++)
212		usleep(1000);
213
214	ptr[256] = 0x1;
215
216	return 0;
217}
218
219static void amdgpu_deadlock_gfx(void)
220{
221	amdgpu_deadlock_helper(AMDGPU_HW_IP_GFX);
222}
223
224static void amdgpu_deadlock_compute(void)
225{
226	amdgpu_deadlock_helper(AMDGPU_HW_IP_COMPUTE);
227}
228
229static void amdgpu_deadlock_helper(unsigned ip_type)
230{
231	amdgpu_context_handle context_handle;
232	amdgpu_bo_handle ib_result_handle;
233	void *ib_result_cpu;
234	uint64_t ib_result_mc_address;
235	struct amdgpu_cs_request ibs_request;
236	struct amdgpu_cs_ib_info ib_info;
237	struct amdgpu_cs_fence fence_status;
238	uint32_t expired;
239	int i, r;
240	amdgpu_bo_list_handle bo_list;
241	amdgpu_va_handle va_handle;
242
243	r = pthread_create(&stress_thread, NULL, write_mem_address, NULL);
244	CU_ASSERT_EQUAL(r, 0);
245
246	r = amdgpu_cs_ctx_create(device_handle, &context_handle);
247	CU_ASSERT_EQUAL(r, 0);
248
249	r = amdgpu_bo_alloc_and_map_raw(device_handle, 4096, 4096,
250			AMDGPU_GEM_DOMAIN_GTT, 0, use_uc_mtype ? AMDGPU_VM_MTYPE_UC : 0,
251						    &ib_result_handle, &ib_result_cpu,
252						    &ib_result_mc_address, &va_handle);
253	CU_ASSERT_EQUAL(r, 0);
254
255	r = amdgpu_get_bo_list(device_handle, ib_result_handle, NULL,
256			       &bo_list);
257	CU_ASSERT_EQUAL(r, 0);
258
259	ptr = ib_result_cpu;
260
261	ptr[0] = PACKET3(PACKET3_WAIT_REG_MEM, 5);
262	ptr[1] = (WAIT_REG_MEM_MEM_SPACE(1) | /* memory */
263			 WAIT_REG_MEM_FUNCTION(4) | /* != */
264			 WAIT_REG_MEM_ENGINE(0));  /* me */
265	ptr[2] = (ib_result_mc_address + 256*4) & 0xfffffffc;
266	ptr[3] = ((ib_result_mc_address + 256*4) >> 32) & 0xffffffff;
267	ptr[4] = 0x00000000; /* reference value */
268	ptr[5] = 0xffffffff; /* and mask */
269	ptr[6] = 0x00000004; /* poll interval */
270
271	for (i = 7; i < 16; ++i)
272		ptr[i] = 0xffff1000;
273
274
275	ptr[256] = 0x0; /* the memory we wait on to change */
276
277
278
279	memset(&ib_info, 0, sizeof(struct amdgpu_cs_ib_info));
280	ib_info.ib_mc_address = ib_result_mc_address;
281	ib_info.size = 16;
282
283	memset(&ibs_request, 0, sizeof(struct amdgpu_cs_request));
284	ibs_request.ip_type = ip_type;
285	ibs_request.ring = 0;
286	ibs_request.number_of_ibs = 1;
287	ibs_request.ibs = &ib_info;
288	ibs_request.resources = bo_list;
289	ibs_request.fence_info.handle = NULL;
290	for (i = 0; i < 200; i++) {
291		r = amdgpu_cs_submit(context_handle, 0,&ibs_request, 1);
292		CU_ASSERT_EQUAL((r == 0 || r == -ECANCELED), 1);
293
294	}
295
296	memset(&fence_status, 0, sizeof(struct amdgpu_cs_fence));
297	fence_status.context = context_handle;
298	fence_status.ip_type = ip_type;
299	fence_status.ip_instance = 0;
300	fence_status.ring = 0;
301	fence_status.fence = ibs_request.seq_no;
302
303	r = amdgpu_cs_query_fence_status(&fence_status,
304			AMDGPU_TIMEOUT_INFINITE,0, &expired);
305	CU_ASSERT_EQUAL((r == 0 || r == -ECANCELED), 1);
306
307	pthread_join(stress_thread, NULL);
308
309	r = amdgpu_bo_list_destroy(bo_list);
310	CU_ASSERT_EQUAL(r, 0);
311
312	r = amdgpu_bo_unmap_and_free(ib_result_handle, va_handle,
313				     ib_result_mc_address, 4096);
314	CU_ASSERT_EQUAL(r, 0);
315
316	r = amdgpu_cs_ctx_free(context_handle);
317	CU_ASSERT_EQUAL(r, 0);
318}
319
320static void amdgpu_deadlock_sdma(void)
321{
322	amdgpu_context_handle context_handle;
323	amdgpu_bo_handle ib_result_handle;
324	void *ib_result_cpu;
325	uint64_t ib_result_mc_address;
326	struct amdgpu_cs_request ibs_request;
327	struct amdgpu_cs_ib_info ib_info;
328	struct amdgpu_cs_fence fence_status;
329	uint32_t expired;
330	int i, r;
331	amdgpu_bo_list_handle bo_list;
332	amdgpu_va_handle va_handle;
333	struct drm_amdgpu_info_hw_ip info;
334	uint32_t ring_id;
335
336	r = amdgpu_query_hw_ip_info(device_handle, AMDGPU_HW_IP_DMA, 0, &info);
337	CU_ASSERT_EQUAL(r, 0);
338
339	r = amdgpu_cs_ctx_create(device_handle, &context_handle);
340	CU_ASSERT_EQUAL(r, 0);
341
342	for (ring_id = 0; (1 << ring_id) & info.available_rings; ring_id++) {
343		r = pthread_create(&stress_thread, NULL, write_mem_address, NULL);
344		CU_ASSERT_EQUAL(r, 0);
345
346		r = amdgpu_bo_alloc_and_map_raw(device_handle, 4096, 4096,
347				AMDGPU_GEM_DOMAIN_GTT, 0, use_uc_mtype ? AMDGPU_VM_MTYPE_UC : 0,
348							    &ib_result_handle, &ib_result_cpu,
349							    &ib_result_mc_address, &va_handle);
350		CU_ASSERT_EQUAL(r, 0);
351
352		r = amdgpu_get_bo_list(device_handle, ib_result_handle, NULL,
353				       &bo_list);
354		CU_ASSERT_EQUAL(r, 0);
355
356		ptr = ib_result_cpu;
357		i = 0;
358
359		ptr[i++] = SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
360				(0 << 26) | /* WAIT_REG_MEM */
361				(4 << 28) | /* != */
362				(1 << 31); /* memory */
363		ptr[i++] = (ib_result_mc_address + 256*4) & 0xfffffffc;
364		ptr[i++] = ((ib_result_mc_address + 256*4) >> 32) & 0xffffffff;
365		ptr[i++] = 0x00000000; /* reference value */
366		ptr[i++] = 0xffffffff; /* and mask */
367		ptr[i++] =  4 | /* poll interval */
368				(0xfff << 16); /* retry count */
369
370		for (; i < 16; i++)
371			ptr[i] = 0;
372
373		ptr[256] = 0x0; /* the memory we wait on to change */
374
375		memset(&ib_info, 0, sizeof(struct amdgpu_cs_ib_info));
376		ib_info.ib_mc_address = ib_result_mc_address;
377		ib_info.size = 16;
378
379		memset(&ibs_request, 0, sizeof(struct amdgpu_cs_request));
380		ibs_request.ip_type = AMDGPU_HW_IP_DMA;
381		ibs_request.ring = ring_id;
382		ibs_request.number_of_ibs = 1;
383		ibs_request.ibs = &ib_info;
384		ibs_request.resources = bo_list;
385		ibs_request.fence_info.handle = NULL;
386
387		for (i = 0; i < 200; i++) {
388			r = amdgpu_cs_submit(context_handle, 0,&ibs_request, 1);
389			CU_ASSERT_EQUAL((r == 0 || r == -ECANCELED), 1);
390
391		}
392
393		memset(&fence_status, 0, sizeof(struct amdgpu_cs_fence));
394		fence_status.context = context_handle;
395		fence_status.ip_type = AMDGPU_HW_IP_DMA;
396		fence_status.ip_instance = 0;
397		fence_status.ring = ring_id;
398		fence_status.fence = ibs_request.seq_no;
399
400		r = amdgpu_cs_query_fence_status(&fence_status,
401				AMDGPU_TIMEOUT_INFINITE,0, &expired);
402		CU_ASSERT_EQUAL((r == 0 || r == -ECANCELED), 1);
403
404		pthread_join(stress_thread, NULL);
405
406		r = amdgpu_bo_list_destroy(bo_list);
407		CU_ASSERT_EQUAL(r, 0);
408
409		r = amdgpu_bo_unmap_and_free(ib_result_handle, va_handle,
410					     ib_result_mc_address, 4096);
411		CU_ASSERT_EQUAL(r, 0);
412	}
413	r = amdgpu_cs_ctx_free(context_handle);
414	CU_ASSERT_EQUAL(r, 0);
415}
416
417static void bad_access_helper(int reg_access)
418{
419	amdgpu_context_handle context_handle;
420	amdgpu_bo_handle ib_result_handle;
421	void *ib_result_cpu;
422	uint64_t ib_result_mc_address;
423	struct amdgpu_cs_request ibs_request;
424	struct amdgpu_cs_ib_info ib_info;
425	struct amdgpu_cs_fence fence_status;
426	uint32_t expired;
427	int i, r;
428	amdgpu_bo_list_handle bo_list;
429	amdgpu_va_handle va_handle;
430
431	r = amdgpu_cs_ctx_create(device_handle, &context_handle);
432	CU_ASSERT_EQUAL(r, 0);
433
434	r = amdgpu_bo_alloc_and_map_raw(device_handle, 4096, 4096,
435			AMDGPU_GEM_DOMAIN_GTT, 0, 0,
436							&ib_result_handle, &ib_result_cpu,
437							&ib_result_mc_address, &va_handle);
438	CU_ASSERT_EQUAL(r, 0);
439
440	r = amdgpu_get_bo_list(device_handle, ib_result_handle, NULL,
441				   &bo_list);
442	CU_ASSERT_EQUAL(r, 0);
443
444	ptr = ib_result_cpu;
445	i = 0;
446
447	ptr[i++] = PACKET3(PACKET3_WRITE_DATA, 3);
448	ptr[i++] = (reg_access ? WRITE_DATA_DST_SEL(0) : WRITE_DATA_DST_SEL(5))| WR_CONFIRM;
449	ptr[i++] = reg_access ? mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR : 0xdeadbee0;
450	ptr[i++] = 0;
451	ptr[i++] = 0xdeadbeef;
452
453	for (; i < 16; ++i)
454		ptr[i] = 0xffff1000;
455
456	memset(&ib_info, 0, sizeof(struct amdgpu_cs_ib_info));
457	ib_info.ib_mc_address = ib_result_mc_address;
458	ib_info.size = 16;
459
460	memset(&ibs_request, 0, sizeof(struct amdgpu_cs_request));
461	ibs_request.ip_type = AMDGPU_HW_IP_GFX;
462	ibs_request.ring = 0;
463	ibs_request.number_of_ibs = 1;
464	ibs_request.ibs = &ib_info;
465	ibs_request.resources = bo_list;
466	ibs_request.fence_info.handle = NULL;
467
468	r = amdgpu_cs_submit(context_handle, 0,&ibs_request, 1);
469	CU_ASSERT_EQUAL((r == 0 || r == -ECANCELED), 1);
470
471
472	memset(&fence_status, 0, sizeof(struct amdgpu_cs_fence));
473	fence_status.context = context_handle;
474	fence_status.ip_type = AMDGPU_HW_IP_GFX;
475	fence_status.ip_instance = 0;
476	fence_status.ring = 0;
477	fence_status.fence = ibs_request.seq_no;
478
479	r = amdgpu_cs_query_fence_status(&fence_status,
480			AMDGPU_TIMEOUT_INFINITE,0, &expired);
481	CU_ASSERT_EQUAL((r == 0 || r == -ECANCELED), 1);
482
483	r = amdgpu_bo_list_destroy(bo_list);
484	CU_ASSERT_EQUAL(r, 0);
485
486	r = amdgpu_bo_unmap_and_free(ib_result_handle, va_handle,
487					 ib_result_mc_address, 4096);
488	CU_ASSERT_EQUAL(r, 0);
489
490	r = amdgpu_cs_ctx_free(context_handle);
491	CU_ASSERT_EQUAL(r, 0);
492}
493
494static void amdgpu_illegal_reg_access()
495{
496	bad_access_helper(1);
497}
498
499static void amdgpu_illegal_mem_access()
500{
501	bad_access_helper(0);
502}
503
504static void amdgpu_dispatch_hang_gfx(void)
505{
506	amdgpu_dispatch_hang_helper(device_handle, AMDGPU_HW_IP_GFX);
507}
508
509static void amdgpu_dispatch_hang_compute(void)
510{
511	amdgpu_dispatch_hang_helper(device_handle, AMDGPU_HW_IP_COMPUTE);
512}
513
514static void amdgpu_dispatch_hang_slow_gfx(void)
515{
516	amdgpu_dispatch_hang_slow_helper(device_handle, AMDGPU_HW_IP_GFX);
517}
518
519static void amdgpu_dispatch_hang_slow_compute(void)
520{
521	amdgpu_dispatch_hang_slow_helper(device_handle, AMDGPU_HW_IP_COMPUTE);
522}
523
524static void amdgpu_draw_hang_gfx(void)
525{
526	int r;
527	struct drm_amdgpu_info_hw_ip info;
528	uint32_t ring_id;
529
530	r = amdgpu_query_hw_ip_info(device_handle, AMDGPU_HW_IP_GFX, 0, &info);
531	CU_ASSERT_EQUAL(r, 0);
532	if (!info.available_rings)
533		printf("SKIP ... as there's no graphic ring\n");
534
535	for (ring_id = 0; (1 << ring_id) & info.available_rings; ring_id++) {
536		amdgpu_memcpy_draw_test(device_handle, ring_id, 0);
537		amdgpu_memcpy_draw_test(device_handle, ring_id, 1);
538		amdgpu_memcpy_draw_test(device_handle, ring_id, 0);
539	}
540}
541
542static void amdgpu_draw_hang_slow_gfx(void)
543{
544	struct drm_amdgpu_info_hw_ip info;
545	uint32_t ring_id;
546	int r;
547
548	r = amdgpu_query_hw_ip_info(device_handle, AMDGPU_HW_IP_GFX, 0, &info);
549	CU_ASSERT_EQUAL(r, 0);
550
551	for (ring_id = 0; (1 << ring_id) & info.available_rings; ring_id++) {
552		amdgpu_memcpy_draw_test(device_handle, ring_id, 0);
553		amdgpu_memcpy_draw_hang_slow_test(device_handle, ring_id);
554		amdgpu_memcpy_draw_test(device_handle, ring_id, 0);
555	}
556}
557