forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2024-01-04 1543e317f1da31b75942316931e8f491a8920811 
 kernel/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
....@@ -22,93 +22,58 @@
2222 
2323 #include "amdgpu_amdkfd.h"
2424 #include "amd_shared.h"
25
-#include <drm/drmP.h>
25
+
2626 #include "amdgpu.h"
2727 #include "amdgpu_gfx.h"
28
+#include "amdgpu_dma_buf.h"
2829 #include <linux/module.h>
30
+#include <linux/dma-buf.h>
31
+#include "amdgpu_xgmi.h"
32
+#include <uapi/linux/kfd_ioctl.h>
2933 
30
-const struct kgd2kfd_calls *kgd2kfd;
31
-bool (*kgd2kfd_init_p)(unsigned int, const struct kgd2kfd_calls**);
34
+/* Total memory size in system memory and all GPU VRAM. Used to
35
+ * estimate worst case amount of memory to reserve for page tables
36
+ */
37
+uint64_t amdgpu_amdkfd_total_mem_size;
3238 
33
-static const unsigned int compute_vmid_bitmap = 0xFF00;
39
+static bool kfd_initialized;
3440 
3541 int amdgpu_amdkfd_init(void)
3642 {
43
+	struct sysinfo si;
3744 	int ret;
3845 
39
-#if defined(CONFIG_HSA_AMD_MODULE)
40
-	int (*kgd2kfd_init_p)(unsigned int, const struct kgd2kfd_calls**);
46
+	si_meminfo(&si);
47
+	amdgpu_amdkfd_total_mem_size = si.totalram - si.totalhigh;
48
+	amdgpu_amdkfd_total_mem_size *= si.mem_unit;
4149 
42
-	kgd2kfd_init_p = symbol_request(kgd2kfd_init);
43
-
44
-	if (kgd2kfd_init_p == NULL)
45
-		return -ENOENT;
46
-
47
-	ret = kgd2kfd_init_p(KFD_INTERFACE_VERSION, &kgd2kfd);
48
-	if (ret) {
49
-		symbol_put(kgd2kfd_init);
50
-		kgd2kfd = NULL;
51
-	}
52
-
53
-
54
-#elif defined(CONFIG_HSA_AMD)
55
-
56
-	ret = kgd2kfd_init(KFD_INTERFACE_VERSION, &kgd2kfd);
57
-	if (ret)
58
-		kgd2kfd = NULL;
59
-
60
-#else
61
-	kgd2kfd = NULL;
62
-	ret = -ENOENT;
63
-#endif
64
-
65
-#if defined(CONFIG_HSA_AMD_MODULE) || defined(CONFIG_HSA_AMD)
50
+	ret = kgd2kfd_init();
6651 	amdgpu_amdkfd_gpuvm_init_mem_limits();
67
-#endif
52
+	kfd_initialized = !ret;
6853 
6954 	return ret;
7055 }
7156 
7257 void amdgpu_amdkfd_fini(void)
7358 {
74
-	if (kgd2kfd) {
75
-		kgd2kfd->exit();
76
-		symbol_put(kgd2kfd_init);
59
+	if (kfd_initialized) {
60
+		kgd2kfd_exit();
61
+		kfd_initialized = false;
7762 	}
7863 }
7964 
8065 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
8166 {
82
-	const struct kfd2kgd_calls *kfd2kgd;
67
+	bool vf = amdgpu_sriov_vf(adev);
8368 
84
-	if (!kgd2kfd)
69
+	if (!kfd_initialized)
8570 		return;
8671 
87
-	switch (adev->asic_type) {
88
-#ifdef CONFIG_DRM_AMDGPU_CIK
89
-	case CHIP_KAVERI:
90
-	case CHIP_HAWAII:
91
-		kfd2kgd = amdgpu_amdkfd_gfx_7_get_functions();
92
-		break;
93
-#endif
94
-	case CHIP_CARRIZO:
95
-	case CHIP_TONGA:
96
-	case CHIP_FIJI:
97
-	case CHIP_POLARIS10:
98
-	case CHIP_POLARIS11:
99
-		kfd2kgd = amdgpu_amdkfd_gfx_8_0_get_functions();
100
-		break;
101
-	case CHIP_VEGA10:
102
-	case CHIP_RAVEN:
103
-		kfd2kgd = amdgpu_amdkfd_gfx_9_0_get_functions();
104
-		break;
105
-	default:
106
-		dev_info(adev->dev, "kfd not supported on this ASIC\n");
107
-		return;
108
-	}
72
+	adev->kfd.dev = kgd2kfd_probe((struct kgd_dev *)adev,
73
+				      adev->pdev, adev->asic_type, vf);
10974 
110
-	adev->kfd = kgd2kfd->probe((struct kgd_dev *)adev,
111
-				   adev->pdev, kfd2kgd);
75
+	if (adev->kfd.dev)
76
+		amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
11277 }
11378 
11479 /**
....@@ -148,97 +113,99 @@
148113 {
149114 	int i;
150115 	int last_valid_bit;
151
-	if (adev->kfd) {
116
+
117
+	if (adev->kfd.dev) {
152118 		struct kgd2kfd_shared_resources gpu_resources = {
153
-			.compute_vmid_bitmap = compute_vmid_bitmap,
119
+			.compute_vmid_bitmap =
120
+				((1 << AMDGPU_NUM_VMID) - 1) -
121
+				((1 << adev->vm_manager.first_kfd_vmid) - 1),
154122 			.num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
155123 			.num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe,
156124 			.gpuvm_size = min(adev->vm_manager.max_pfn
157125 					  << AMDGPU_GPU_PAGE_SHIFT,
158
-					  AMDGPU_VA_HOLE_START),
159
-			.drm_render_minor = adev->ddev->render->index
126
+					  AMDGPU_GMC_HOLE_START),
127
+			.drm_render_minor = adev_to_drm(adev)->render->index,
128
+			.sdma_doorbell_idx = adev->doorbell_index.sdma_engine,
129
+
160130 		};
161131 
162132 		/* this is going to have a few of the MSBs set that we need to
163
-		 * clear */
164
-		bitmap_complement(gpu_resources.queue_bitmap,
133
+		 * clear
134
+		 */
135
+		bitmap_complement(gpu_resources.cp_queue_bitmap,
165136 				  adev->gfx.mec.queue_bitmap,
166137 				  KGD_MAX_QUEUES);
167138 
168
-		/* remove the KIQ bit as well */
169
-		if (adev->gfx.kiq.ring.ready)
170
-			clear_bit(amdgpu_gfx_queue_to_bit(adev,
171
-							  adev->gfx.kiq.ring.me - 1,
172
-							  adev->gfx.kiq.ring.pipe,
173
-							  adev->gfx.kiq.ring.queue),
174
-				  gpu_resources.queue_bitmap);
175
-
176139 		/* According to linux/bitmap.h we shouldn't use bitmap_clear if
177
-		 * nbits is not compile time constant */
140
+		 * nbits is not compile time constant
141
+		 */
178142 		last_valid_bit = 1 /* only first MEC can have compute queues */
179143 				* adev->gfx.mec.num_pipe_per_mec
180144 				* adev->gfx.mec.num_queue_per_pipe;
181145 		for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i)
182
-			clear_bit(i, gpu_resources.queue_bitmap);
146
+			clear_bit(i, gpu_resources.cp_queue_bitmap);
183147 
184148 		amdgpu_doorbell_get_kfd_info(adev,
185149 				&gpu_resources.doorbell_physical_address,
186150 				&gpu_resources.doorbell_aperture_size,
187151 				&gpu_resources.doorbell_start_offset);
152
+
153
+		/* Since SOC15, BIF starts to statically use the
154
+		 * lower 12 bits of doorbell addresses for routing
155
+		 * based on settings in registers like
156
+		 * SDMA0_DOORBELL_RANGE etc..
157
+		 * In order to route a doorbell to CP engine, the lower
158
+		 * 12 bits of its address has to be outside the range
159
+		 * set for SDMA, VCN, and IH blocks.
160
+		 */
188161 		if (adev->asic_type >= CHIP_VEGA10) {
189
-			/* On SOC15 the BIF is involved in routing
190
-			 * doorbells using the low 12 bits of the
191
-			 * address. Communicate the assignments to
192
-			 * KFD. KFD uses two doorbell pages per
193
-			 * process in case of 64-bit doorbells so we
194
-			 * can use each doorbell assignment twice.
195
-			 */
196
-			gpu_resources.sdma_doorbell[0][0] =
197
-				AMDGPU_DOORBELL64_sDMA_ENGINE0;
198
-			gpu_resources.sdma_doorbell[0][1] =
199
-				AMDGPU_DOORBELL64_sDMA_ENGINE0 + 0x200;
200
-			gpu_resources.sdma_doorbell[1][0] =
201
-				AMDGPU_DOORBELL64_sDMA_ENGINE1;
202
-			gpu_resources.sdma_doorbell[1][1] =
203
-				AMDGPU_DOORBELL64_sDMA_ENGINE1 + 0x200;
204
-			/* Doorbells 0x0f0-0ff and 0x2f0-2ff are reserved for
205
-			 * SDMA, IH and VCN. So don't use them for the CP.
206
-			 */
207
-			gpu_resources.reserved_doorbell_mask = 0x1f0;
208
-			gpu_resources.reserved_doorbell_val  = 0x0f0;
162
+			gpu_resources.non_cp_doorbells_start =
163
+					adev->doorbell_index.first_non_cp;
164
+			gpu_resources.non_cp_doorbells_end =
165
+					adev->doorbell_index.last_non_cp;
209166 		}
210167 
211
-		kgd2kfd->device_init(adev->kfd, &gpu_resources);
168
+		kgd2kfd_device_init(adev->kfd.dev, adev_to_drm(adev), &gpu_resources);
212169 	}
213170 }
214171 
215172 void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev)
216173 {
217
-	if (adev->kfd) {
218
-		kgd2kfd->device_exit(adev->kfd);
219
-		adev->kfd = NULL;
174
+	if (adev->kfd.dev) {
175
+		kgd2kfd_device_exit(adev->kfd.dev);
176
+		adev->kfd.dev = NULL;
220177 	}
221178 }
222179 
223180 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
224181 		const void *ih_ring_entry)
225182 {
226
-	if (adev->kfd)
227
-		kgd2kfd->interrupt(adev->kfd, ih_ring_entry);
183
+	if (adev->kfd.dev)
184
+		kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
228185 }
229186 
230
-void amdgpu_amdkfd_suspend(struct amdgpu_device *adev)
187
+void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm)
231188 {
232
-	if (adev->kfd)
233
-		kgd2kfd->suspend(adev->kfd);
189
+	if (adev->kfd.dev)
190
+		kgd2kfd_suspend(adev->kfd.dev, run_pm);
234191 }
235192 
236
-int amdgpu_amdkfd_resume(struct amdgpu_device *adev)
193
+int amdgpu_amdkfd_resume_iommu(struct amdgpu_device *adev)
237194 {
238195 	int r = 0;
239196 
240
-	if (adev->kfd)
241
-		r = kgd2kfd->resume(adev->kfd);
197
+	if (adev->kfd.dev)
198
+		r = kgd2kfd_resume_iommu(adev->kfd.dev);
199
+
200
+	return r;
201
+}
202
+
203
+int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm)
204
+{
205
+	int r = 0;
206
+
207
+	if (adev->kfd.dev)
208
+		r = kgd2kfd_resume(adev->kfd.dev, run_pm);
242209 
243210 	return r;
244211 }
....@@ -247,8 +214,8 @@
247214 {
248215 	int r = 0;
249216 
250
-	if (adev->kfd)
251
-		r = kgd2kfd->pre_reset(adev->kfd);
217
+	if (adev->kfd.dev)
218
+		r = kgd2kfd_pre_reset(adev->kfd.dev);
252219 
253220 	return r;
254221 }
....@@ -257,8 +224,8 @@
257224 {
258225 	int r = 0;
259226 
260
-	if (adev->kfd)
261
-		r = kgd2kfd->post_reset(adev->kfd);
227
+	if (adev->kfd.dev)
228
+		r = kgd2kfd_post_reset(adev->kfd.dev);
262229 
263230 	return r;
264231 }
....@@ -267,12 +234,13 @@
267234 {
268235 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
269236 
270
-	amdgpu_device_gpu_recover(adev, NULL, false);
237
+	if (amdgpu_device_should_recover_gpu(adev))
238
+		amdgpu_device_gpu_recover(adev, NULL);
271239 }
272240 
273
-int alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
274
-			void **mem_obj, uint64_t *gpu_addr,
275
-			void **cpu_ptr, bool mqd_gfx9)
241
+int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
242
+				void **mem_obj, uint64_t *gpu_addr,
243
+				void **cpu_ptr, bool cp_mqd_gfx9)
276244 {
277245 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
278246 	struct amdgpu_bo *bo = NULL;
....@@ -288,8 +256,8 @@
288256 	bp.type = ttm_bo_type_kernel;
289257 	bp.resv = NULL;
290258 
291
-	if (mqd_gfx9)
292
-		bp.flags |= AMDGPU_GEM_CREATE_MQD_GFX9;
259
+	if (cp_mqd_gfx9)
260
+		bp.flags |= AMDGPU_GEM_CREATE_CP_MQD_GFX9;
293261 
294262 	r = amdgpu_bo_create(adev, &bp, &bo);
295263 	if (r) {
....@@ -342,7 +310,7 @@
342310 	return r;
343311 }
344312 
345
-void free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
313
+void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
346314 {
347315 	struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
348316 
....@@ -353,8 +321,79 @@
353321 	amdgpu_bo_unref(&(bo));
354322 }
355323 
356
-void get_local_mem_info(struct kgd_dev *kgd,
357
-			struct kfd_local_mem_info *mem_info)
324
+int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size,
325
+				void **mem_obj)
326
+{
327
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
328
+	struct amdgpu_bo *bo = NULL;
329
+	struct amdgpu_bo_param bp;
330
+	int r;
331
+
332
+	memset(&bp, 0, sizeof(bp));
333
+	bp.size = size;
334
+	bp.byte_align = 1;
335
+	bp.domain = AMDGPU_GEM_DOMAIN_GWS;
336
+	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
337
+	bp.type = ttm_bo_type_device;
338
+	bp.resv = NULL;
339
+
340
+	r = amdgpu_bo_create(adev, &bp, &bo);
341
+	if (r) {
342
+		dev_err(adev->dev,
343
+			"failed to allocate gws BO for amdkfd (%d)\n", r);
344
+		return r;
345
+	}
346
+
347
+	*mem_obj = bo;
348
+	return 0;
349
+}
350
+
351
+void amdgpu_amdkfd_free_gws(struct kgd_dev *kgd, void *mem_obj)
352
+{
353
+	struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj;
354
+
355
+	amdgpu_bo_unref(&bo);
356
+}
357
+
358
+uint32_t amdgpu_amdkfd_get_fw_version(struct kgd_dev *kgd,
359
+				      enum kgd_engine_type type)
360
+{
361
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
362
+
363
+	switch (type) {
364
+	case KGD_ENGINE_PFP:
365
+		return adev->gfx.pfp_fw_version;
366
+
367
+	case KGD_ENGINE_ME:
368
+		return adev->gfx.me_fw_version;
369
+
370
+	case KGD_ENGINE_CE:
371
+		return adev->gfx.ce_fw_version;
372
+
373
+	case KGD_ENGINE_MEC1:
374
+		return adev->gfx.mec_fw_version;
375
+
376
+	case KGD_ENGINE_MEC2:
377
+		return adev->gfx.mec2_fw_version;
378
+
379
+	case KGD_ENGINE_RLC:
380
+		return adev->gfx.rlc_fw_version;
381
+
382
+	case KGD_ENGINE_SDMA1:
383
+		return adev->sdma.instance[0].fw_version;
384
+
385
+	case KGD_ENGINE_SDMA2:
386
+		return adev->sdma.instance[1].fw_version;
387
+
388
+	default:
389
+		return 0;
390
+	}
391
+
392
+	return 0;
393
+}
394
+
395
+void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd,
396
+				      struct kfd_local_mem_info *mem_info)
358397 {
359398 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
360399 	uint64_t address_mask = adev->dev->dma_mask ? ~*adev->dev->dma_mask :
....@@ -379,13 +418,16 @@
379418 
380419 	if (amdgpu_sriov_vf(adev))
381420 		mem_info->mem_clk_max = adev->clock.default_mclk / 100;
382
-	else if (adev->powerplay.pp_funcs)
383
-		mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
384
-	else
421
+	else if (adev->pm.dpm_enabled) {
422
+		if (amdgpu_emu_mode == 1)
423
+			mem_info->mem_clk_max = 0;
424
+		else
425
+			mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
426
+	} else
385427 		mem_info->mem_clk_max = 100;
386428 }
387429 
388
-uint64_t get_gpu_clock_counter(struct kgd_dev *kgd)
430
+uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd)
389431 {
390432 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
391433 
....@@ -394,20 +436,20 @@
394436 	return 0;
395437 }
396438 
397
-uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
439
+uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
398440 {
399441 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
400442 
401443 	/* the sclk is in quantas of 10kHz */
402444 	if (amdgpu_sriov_vf(adev))
403445 		return adev->clock.default_sclk / 100;
404
-	else if (adev->powerplay.pp_funcs)
446
+	else if (adev->pm.dpm_enabled)
405447 		return amdgpu_dpm_get_sclk(adev, false) / 100;
406448 	else
407449 		return 100;
408450 }
409451 
410
-void get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
452
+void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
411453 {
412454 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
413455 	struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
....@@ -430,11 +472,126 @@
430472 	cu_info->lds_size = acu_info.lds_size;
431473 }
432474 
475
+int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,
476
+				  struct kgd_dev **dma_buf_kgd,
477
+				  uint64_t *bo_size, void *metadata_buffer,
478
+				  size_t buffer_size, uint32_t *metadata_size,
479
+				  uint32_t *flags)
480
+{
481
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
482
+	struct dma_buf *dma_buf;
483
+	struct drm_gem_object *obj;
484
+	struct amdgpu_bo *bo;
485
+	uint64_t metadata_flags;
486
+	int r = -EINVAL;
487
+
488
+	dma_buf = dma_buf_get(dma_buf_fd);
489
+	if (IS_ERR(dma_buf))
490
+		return PTR_ERR(dma_buf);
491
+
492
+	if (dma_buf->ops != &amdgpu_dmabuf_ops)
493
+		/* Can't handle non-graphics buffers */
494
+		goto out_put;
495
+
496
+	obj = dma_buf->priv;
497
+	if (obj->dev->driver != adev_to_drm(adev)->driver)
498
+		/* Can't handle buffers from different drivers */
499
+		goto out_put;
500
+
501
+	adev = drm_to_adev(obj->dev);
502
+	bo = gem_to_amdgpu_bo(obj);
503
+	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
504
+				    AMDGPU_GEM_DOMAIN_GTT)))
505
+		/* Only VRAM and GTT BOs are supported */
506
+		goto out_put;
507
+
508
+	r = 0;
509
+	if (dma_buf_kgd)
510
+		*dma_buf_kgd = (struct kgd_dev *)adev;
511
+	if (bo_size)
512
+		*bo_size = amdgpu_bo_size(bo);
513
+	if (metadata_size)
514
+		*metadata_size = bo->metadata_size;
515
+	if (metadata_buffer)
516
+		r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
517
+					   metadata_size, &metadata_flags);
518
+	if (flags) {
519
+		*flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
520
+				KFD_IOC_ALLOC_MEM_FLAGS_VRAM
521
+				: KFD_IOC_ALLOC_MEM_FLAGS_GTT;
522
+
523
+		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
524
+			*flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC;
525
+	}
526
+
527
+out_put:
528
+	dma_buf_put(dma_buf);
529
+	return r;
530
+}
531
+
433532 uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd)
434533 {
435534 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
535
+	struct ttm_resource_manager *vram_man = ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM);
436536 
437
-	return amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
537
+	return amdgpu_vram_mgr_usage(vram_man);
538
+}
539
+
540
+uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd)
541
+{
542
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
543
+
544
+	return adev->gmc.xgmi.hive_id;
545
+}
546
+
547
+uint64_t amdgpu_amdkfd_get_unique_id(struct kgd_dev *kgd)
548
+{
549
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
550
+
551
+	return adev->unique_id;
552
+}
553
+
554
+uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *src)
555
+{
556
+	struct amdgpu_device *peer_adev = (struct amdgpu_device *)src;
557
+	struct amdgpu_device *adev = (struct amdgpu_device *)dst;
558
+	int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev);
559
+
560
+	if (ret < 0) {
561
+		DRM_ERROR("amdgpu: failed to get  xgmi hops count between node %d and %d. ret = %d\n",
562
+			adev->gmc.xgmi.physical_node_id,
563
+			peer_adev->gmc.xgmi.physical_node_id, ret);
564
+		ret = 0;
565
+	}
566
+	return  (uint8_t)ret;
567
+}
568
+
569
+uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd)
570
+{
571
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
572
+
573
+	return adev->rmmio_remap.bus_addr;
574
+}
575
+
576
+uint32_t amdgpu_amdkfd_get_num_gws(struct kgd_dev *kgd)
577
+{
578
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
579
+
580
+	return adev->gds.gws_size;
581
+}
582
+
583
+uint32_t amdgpu_amdkfd_get_asic_rev_id(struct kgd_dev *kgd)
584
+{
585
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
586
+
587
+	return adev->rev_id;
588
+}
589
+
590
+int amdgpu_amdkfd_get_noretry(struct kgd_dev *kgd)
591
+{
592
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
593
+
594
+	return adev->gmc.noretry;
438595 }
439596 
440597 int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
....@@ -478,6 +635,7 @@
478635 	job->vmid = vmid;
479636 
480637 	ret = amdgpu_ib_schedule(ring, 1, ib, job, &f);
638
+
481639 	if (ret) {
482640 		DRM_ERROR("amdgpu: failed to schedule IB.\n");
483641 		goto err_ib_sched;
....@@ -496,60 +654,51 @@
496654 {
497655 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
498656 
499
-	if (adev->powerplay.pp_funcs &&
500
-	    adev->powerplay.pp_funcs->switch_power_profile)
501
-		amdgpu_dpm_switch_power_profile(adev,
502
-						PP_SMC_POWER_PROFILE_COMPUTE,
503
-						!idle);
657
+	amdgpu_dpm_switch_power_profile(adev,
658
+					PP_SMC_POWER_PROFILE_COMPUTE,
659
+					!idle);
504660 }
505661 
506662 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
507663 {
508
-	if (adev->kfd) {
509
-		if ((1 << vmid) & compute_vmid_bitmap)
510
-			return true;
664
+	if (adev->kfd.dev)
665
+		return vmid >= adev->vm_manager.first_kfd_vmid;
666
+
667
+	return false;
668
+}
669
+
670
+int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct kgd_dev *kgd, uint16_t vmid)
671
+{
672
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
673
+
674
+	if (adev->family == AMDGPU_FAMILY_AI) {
675
+		int i;
676
+
677
+		for (i = 0; i < adev->num_vmhubs; i++)
678
+			amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
679
+	} else {
680
+		amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0);
511681 	}
512682 
513
-	return false;
514
-}
515
-
516
-#if !defined(CONFIG_HSA_AMD_MODULE) && !defined(CONFIG_HSA_AMD)
517
-bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
518
-{
519
-	return false;
520
-}
521
-
522
-void amdgpu_amdkfd_unreserve_system_memory_limit(struct amdgpu_bo *bo)
523
-{
524
-}
525
-
526
-void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
527
-					struct amdgpu_vm *vm)
528
-{
529
-}
530
-
531
-struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
532
-{
533
-	return NULL;
534
-}
535
-
536
-int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm)
537
-{
538683 	return 0;
539684 }
540685 
541
-struct kfd2kgd_calls *amdgpu_amdkfd_gfx_7_get_functions(void)
686
+int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct kgd_dev *kgd, uint16_t pasid)
542687 {
543
-	return NULL;
688
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
689
+	const uint32_t flush_type = 0;
690
+	bool all_hub = false;
691
+
692
+	if (adev->family == AMDGPU_FAMILY_AI ||
693
+	    adev->family == AMDGPU_FAMILY_RV)
694
+		all_hub = true;
695
+
696
+	return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub);
544697 }
545698 
546
-struct kfd2kgd_calls *amdgpu_amdkfd_gfx_8_0_get_functions(void)
699
+bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd)
547700 {
548
-	return NULL;
549
-}
701
+	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
550702 
551
-struct kfd2kgd_calls *amdgpu_amdkfd_gfx_9_0_get_functions(void)
552
-{
553
-	return NULL;
703
+	return adev->have_atomics_support;
554704 }
555
-#endif