1 /*
2  * Copyright 2014 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 #include "amdgpu_amdkfd.h"
24 #include "amd_pcie.h"
25 #include "amd_shared.h"
26 
27 #include "amdgpu.h"
28 #include "amdgpu_gfx.h"
29 #include "amdgpu_dma_buf.h"
30 #include <linux/module.h>
31 #include <linux/dma-buf.h>
32 #include "amdgpu_xgmi.h"
33 #include <uapi/linux/kfd_ioctl.h>
34 #include "amdgpu_ras.h"
35 #include "amdgpu_umc.h"
36 #include "amdgpu_reset.h"
37 
38 /* Total memory size in system memory and all GPU VRAM. Used to
39  * estimate worst case amount of memory to reserve for page tables
40  */
41 uint64_t amdgpu_amdkfd_total_mem_size;
42 
43 static bool kfd_initialized;
44 
45 int amdgpu_amdkfd_init(void)
46 {
47 	struct sysinfo si;
48 	int ret;
49 
50 	si_meminfo(&si);
51 	amdgpu_amdkfd_total_mem_size = si.freeram - si.freehigh;
52 	amdgpu_amdkfd_total_mem_size *= si.mem_unit;
53 
54 	ret = kgd2kfd_init();
55 	amdgpu_amdkfd_gpuvm_init_mem_limits();
56 	kfd_initialized = !ret;
57 
58 	return ret;
59 }
60 
61 void amdgpu_amdkfd_fini(void)
62 {
63 	if (kfd_initialized) {
64 		kgd2kfd_exit();
65 		kfd_initialized = false;
66 	}
67 }
68 
69 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
70 {
71 	bool vf = amdgpu_sriov_vf(adev);
72 
73 	if (!kfd_initialized)
74 		return;
75 
76 	adev->kfd.dev = kgd2kfd_probe(adev, vf);
77 
78 	if (adev->kfd.dev)
79 		amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
80 }
81 
82 /**
83  * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
84  *                                setup amdkfd
85  *
86  * @adev: amdgpu_device pointer
87  * @aperture_base: output returning doorbell aperture base physical address
88  * @aperture_size: output returning doorbell aperture size in bytes
89  * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
90  *
91  * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
92  * takes doorbells required for its own rings and reports the setup to amdkfd.
93  * amdgpu reserved doorbells are at the start of the doorbell aperture.
94  */
95 static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
96 					 phys_addr_t *aperture_base,
97 					 size_t *aperture_size,
98 					 size_t *start_offset)
99 {
100 	/*
101 	 * The first num_doorbells are used by amdgpu.
102 	 * amdkfd takes whatever's left in the aperture.
103 	 */
104 	if (adev->enable_mes) {
105 		/*
106 		 * With MES enabled, we only need to initialize
107 		 * the base address. The size and offset are
108 		 * not initialized as AMDGPU manages the whole
109 		 * doorbell space.
110 		 */
111 		*aperture_base = adev->doorbell.base;
112 		*aperture_size = 0;
113 		*start_offset = 0;
114 	} else if (adev->doorbell.size > adev->doorbell.num_doorbells *
115 						sizeof(u32)) {
116 		*aperture_base = adev->doorbell.base;
117 		*aperture_size = adev->doorbell.size;
118 		*start_offset = adev->doorbell.num_doorbells * sizeof(u32);
119 	} else {
120 		*aperture_base = 0;
121 		*aperture_size = 0;
122 		*start_offset = 0;
123 	}
124 }
125 
126 
127 static void amdgpu_amdkfd_reset_work(struct work_struct *work)
128 {
129 	struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
130 						  kfd.reset_work);
131 
132 	struct amdgpu_reset_context reset_context;
133 	memset(&reset_context, 0, sizeof(reset_context));
134 
135 	reset_context.method = AMD_RESET_METHOD_NONE;
136 	reset_context.reset_req_dev = adev;
137 	clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
138 
139 	amdgpu_device_gpu_recover(adev, NULL, &reset_context);
140 }
141 
142 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
143 {
144 	int i;
145 	int last_valid_bit;
146 
147 	if (adev->kfd.dev) {
148 		struct kgd2kfd_shared_resources gpu_resources = {
149 			.compute_vmid_bitmap =
150 				((1 << AMDGPU_NUM_VMID) - 1) -
151 				((1 << adev->vm_manager.first_kfd_vmid) - 1),
152 			.num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
153 			.num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe,
154 			.gpuvm_size = min(adev->vm_manager.max_pfn
155 					  << AMDGPU_GPU_PAGE_SHIFT,
156 					  AMDGPU_GMC_HOLE_START),
157 			.drm_render_minor = adev_to_drm(adev)->render->index,
158 			.sdma_doorbell_idx = adev->doorbell_index.sdma_engine,
159 			.enable_mes = adev->enable_mes,
160 		};
161 
162 		/* this is going to have a few of the MSBs set that we need to
163 		 * clear
164 		 */
165 		bitmap_complement(gpu_resources.cp_queue_bitmap,
166 				  adev->gfx.mec.queue_bitmap,
167 				  KGD_MAX_QUEUES);
168 
169 		/* According to linux/bitmap.h we shouldn't use bitmap_clear if
170 		 * nbits is not compile time constant
171 		 */
172 		last_valid_bit = 1 /* only first MEC can have compute queues */
173 				* adev->gfx.mec.num_pipe_per_mec
174 				* adev->gfx.mec.num_queue_per_pipe;
175 		for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i)
176 			clear_bit(i, gpu_resources.cp_queue_bitmap);
177 
178 		amdgpu_doorbell_get_kfd_info(adev,
179 				&gpu_resources.doorbell_physical_address,
180 				&gpu_resources.doorbell_aperture_size,
181 				&gpu_resources.doorbell_start_offset);
182 
183 		/* Since SOC15, BIF starts to statically use the
184 		 * lower 12 bits of doorbell addresses for routing
185 		 * based on settings in registers like
186 		 * SDMA0_DOORBELL_RANGE etc..
187 		 * In order to route a doorbell to CP engine, the lower
188 		 * 12 bits of its address has to be outside the range
189 		 * set for SDMA, VCN, and IH blocks.
190 		 */
191 		if (adev->asic_type >= CHIP_VEGA10) {
192 			gpu_resources.non_cp_doorbells_start =
193 					adev->doorbell_index.first_non_cp;
194 			gpu_resources.non_cp_doorbells_end =
195 					adev->doorbell_index.last_non_cp;
196 		}
197 
198 		adev->kfd.init_complete = kgd2kfd_device_init(adev->kfd.dev,
199 						adev_to_drm(adev), &gpu_resources);
200 
201 		INIT_WORK(&adev->kfd.reset_work, amdgpu_amdkfd_reset_work);
202 	}
203 }
204 
205 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev)
206 {
207 	if (adev->kfd.dev) {
208 		kgd2kfd_device_exit(adev->kfd.dev);
209 		adev->kfd.dev = NULL;
210 	}
211 }
212 
213 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
214 		const void *ih_ring_entry)
215 {
216 	if (adev->kfd.dev)
217 		kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
218 }
219 
220 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm)
221 {
222 	if (adev->kfd.dev)
223 		kgd2kfd_suspend(adev->kfd.dev, run_pm);
224 }
225 
226 int amdgpu_amdkfd_resume_iommu(struct amdgpu_device *adev)
227 {
228 	int r = 0;
229 
230 	if (adev->kfd.dev)
231 		r = kgd2kfd_resume_iommu(adev->kfd.dev);
232 
233 	return r;
234 }
235 
236 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm)
237 {
238 	int r = 0;
239 
240 	if (adev->kfd.dev)
241 		r = kgd2kfd_resume(adev->kfd.dev, run_pm);
242 
243 	return r;
244 }
245 
246 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
247 {
248 	int r = 0;
249 
250 	if (adev->kfd.dev)
251 		r = kgd2kfd_pre_reset(adev->kfd.dev);
252 
253 	return r;
254 }
255 
256 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
257 {
258 	int r = 0;
259 
260 	if (adev->kfd.dev)
261 		r = kgd2kfd_post_reset(adev->kfd.dev);
262 
263 	return r;
264 }
265 
266 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev)
267 {
268 	if (amdgpu_device_should_recover_gpu(adev))
269 		amdgpu_reset_domain_schedule(adev->reset_domain,
270 					     &adev->kfd.reset_work);
271 }
272 
273 int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size,
274 				void **mem_obj, uint64_t *gpu_addr,
275 				void **cpu_ptr, bool cp_mqd_gfx9)
276 {
277 	struct amdgpu_bo *bo = NULL;
278 	struct amdgpu_bo_param bp;
279 	int r;
280 	void *cpu_ptr_tmp = NULL;
281 
282 	memset(&bp, 0, sizeof(bp));
283 	bp.size = size;
284 	bp.byte_align = PAGE_SIZE;
285 	bp.domain = AMDGPU_GEM_DOMAIN_GTT;
286 	bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
287 	bp.type = ttm_bo_type_kernel;
288 	bp.resv = NULL;
289 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
290 
291 	if (cp_mqd_gfx9)
292 		bp.flags |= AMDGPU_GEM_CREATE_CP_MQD_GFX9;
293 
294 	r = amdgpu_bo_create(adev, &bp, &bo);
295 	if (r) {
296 		dev_err(adev->dev,
297 			"failed to allocate BO for amdkfd (%d)\n", r);
298 		return r;
299 	}
300 
301 	/* map the buffer */
302 	r = amdgpu_bo_reserve(bo, true);
303 	if (r) {
304 		dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
305 		goto allocate_mem_reserve_bo_failed;
306 	}
307 
308 	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
309 	if (r) {
310 		dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r);
311 		goto allocate_mem_pin_bo_failed;
312 	}
313 
314 	r = amdgpu_ttm_alloc_gart(&bo->tbo);
315 	if (r) {
316 		dev_err(adev->dev, "%p bind failed\n", bo);
317 		goto allocate_mem_kmap_bo_failed;
318 	}
319 
320 	r = amdgpu_bo_kmap(bo, &cpu_ptr_tmp);
321 	if (r) {
322 		dev_err(adev->dev,
323 			"(%d) failed to map bo to kernel for amdkfd\n", r);
324 		goto allocate_mem_kmap_bo_failed;
325 	}
326 
327 	*mem_obj = bo;
328 	*gpu_addr = amdgpu_bo_gpu_offset(bo);
329 	*cpu_ptr = cpu_ptr_tmp;
330 
331 	amdgpu_bo_unreserve(bo);
332 
333 	return 0;
334 
335 allocate_mem_kmap_bo_failed:
336 	amdgpu_bo_unpin(bo);
337 allocate_mem_pin_bo_failed:
338 	amdgpu_bo_unreserve(bo);
339 allocate_mem_reserve_bo_failed:
340 	amdgpu_bo_unref(&bo);
341 
342 	return r;
343 }
344 
345 void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void *mem_obj)
346 {
347 	struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
348 
349 	amdgpu_bo_reserve(bo, true);
350 	amdgpu_bo_kunmap(bo);
351 	amdgpu_bo_unpin(bo);
352 	amdgpu_bo_unreserve(bo);
353 	amdgpu_bo_unref(&(bo));
354 }
355 
356 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size,
357 				void **mem_obj)
358 {
359 	struct amdgpu_bo *bo = NULL;
360 	struct amdgpu_bo_user *ubo;
361 	struct amdgpu_bo_param bp;
362 	int r;
363 
364 	memset(&bp, 0, sizeof(bp));
365 	bp.size = size;
366 	bp.byte_align = 1;
367 	bp.domain = AMDGPU_GEM_DOMAIN_GWS;
368 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
369 	bp.type = ttm_bo_type_device;
370 	bp.resv = NULL;
371 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
372 
373 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
374 	if (r) {
375 		dev_err(adev->dev,
376 			"failed to allocate gws BO for amdkfd (%d)\n", r);
377 		return r;
378 	}
379 
380 	bo = &ubo->bo;
381 	*mem_obj = bo;
382 	return 0;
383 }
384 
385 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj)
386 {
387 	struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj;
388 
389 	amdgpu_bo_unref(&bo);
390 }
391 
392 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev,
393 				      enum kgd_engine_type type)
394 {
395 	switch (type) {
396 	case KGD_ENGINE_PFP:
397 		return adev->gfx.pfp_fw_version;
398 
399 	case KGD_ENGINE_ME:
400 		return adev->gfx.me_fw_version;
401 
402 	case KGD_ENGINE_CE:
403 		return adev->gfx.ce_fw_version;
404 
405 	case KGD_ENGINE_MEC1:
406 		return adev->gfx.mec_fw_version;
407 
408 	case KGD_ENGINE_MEC2:
409 		return adev->gfx.mec2_fw_version;
410 
411 	case KGD_ENGINE_RLC:
412 		return adev->gfx.rlc_fw_version;
413 
414 	case KGD_ENGINE_SDMA1:
415 		return adev->sdma.instance[0].fw_version;
416 
417 	case KGD_ENGINE_SDMA2:
418 		return adev->sdma.instance[1].fw_version;
419 
420 	default:
421 		return 0;
422 	}
423 
424 	return 0;
425 }
426 
427 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
428 				      struct kfd_local_mem_info *mem_info)
429 {
430 	memset(mem_info, 0, sizeof(*mem_info));
431 
432 	mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
433 	mem_info->local_mem_size_private = adev->gmc.real_vram_size -
434 						adev->gmc.visible_vram_size;
435 
436 	mem_info->vram_width = adev->gmc.vram_width;
437 
438 	pr_debug("Address base: %pap public 0x%llx private 0x%llx\n",
439 			&adev->gmc.aper_base,
440 			mem_info->local_mem_size_public,
441 			mem_info->local_mem_size_private);
442 
443 	if (amdgpu_sriov_vf(adev))
444 		mem_info->mem_clk_max = adev->clock.default_mclk / 100;
445 	else if (adev->pm.dpm_enabled) {
446 		if (amdgpu_emu_mode == 1)
447 			mem_info->mem_clk_max = 0;
448 		else
449 			mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
450 	} else
451 		mem_info->mem_clk_max = 100;
452 }
453 
454 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev)
455 {
456 	if (adev->gfx.funcs->get_gpu_clock_counter)
457 		return adev->gfx.funcs->get_gpu_clock_counter(adev);
458 	return 0;
459 }
460 
461 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev)
462 {
463 	/* the sclk is in quantas of 10kHz */
464 	if (amdgpu_sriov_vf(adev))
465 		return adev->clock.default_sclk / 100;
466 	else if (adev->pm.dpm_enabled)
467 		return amdgpu_dpm_get_sclk(adev, false) / 100;
468 	else
469 		return 100;
470 }
471 
472 void amdgpu_amdkfd_get_cu_info(struct amdgpu_device *adev, struct kfd_cu_info *cu_info)
473 {
474 	struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
475 
476 	memset(cu_info, 0, sizeof(*cu_info));
477 	if (sizeof(cu_info->cu_bitmap) != sizeof(acu_info.bitmap))
478 		return;
479 
480 	cu_info->cu_active_number = acu_info.number;
481 	cu_info->cu_ao_mask = acu_info.ao_cu_mask;
482 	memcpy(&cu_info->cu_bitmap[0], &acu_info.bitmap[0],
483 	       sizeof(acu_info.bitmap));
484 	cu_info->num_shader_engines = adev->gfx.config.max_shader_engines;
485 	cu_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se;
486 	cu_info->num_cu_per_sh = adev->gfx.config.max_cu_per_sh;
487 	cu_info->simd_per_cu = acu_info.simd_per_cu;
488 	cu_info->max_waves_per_simd = acu_info.max_waves_per_simd;
489 	cu_info->wave_front_size = acu_info.wave_front_size;
490 	cu_info->max_scratch_slots_per_cu = acu_info.max_scratch_slots_per_cu;
491 	cu_info->lds_size = acu_info.lds_size;
492 }
493 
494 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd,
495 				  struct amdgpu_device **dmabuf_adev,
496 				  uint64_t *bo_size, void *metadata_buffer,
497 				  size_t buffer_size, uint32_t *metadata_size,
498 				  uint32_t *flags)
499 {
500 	struct dma_buf *dma_buf;
501 	struct drm_gem_object *obj;
502 	struct amdgpu_bo *bo;
503 	uint64_t metadata_flags;
504 	int r = -EINVAL;
505 
506 	dma_buf = dma_buf_get(dma_buf_fd);
507 	if (IS_ERR(dma_buf))
508 		return PTR_ERR(dma_buf);
509 
510 	if (dma_buf->ops != &amdgpu_dmabuf_ops)
511 		/* Can't handle non-graphics buffers */
512 		goto out_put;
513 
514 	obj = dma_buf->priv;
515 	if (obj->dev->driver != adev_to_drm(adev)->driver)
516 		/* Can't handle buffers from different drivers */
517 		goto out_put;
518 
519 	adev = drm_to_adev(obj->dev);
520 	bo = gem_to_amdgpu_bo(obj);
521 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
522 				    AMDGPU_GEM_DOMAIN_GTT)))
523 		/* Only VRAM and GTT BOs are supported */
524 		goto out_put;
525 
526 	r = 0;
527 	if (dmabuf_adev)
528 		*dmabuf_adev = adev;
529 	if (bo_size)
530 		*bo_size = amdgpu_bo_size(bo);
531 	if (metadata_buffer)
532 		r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
533 					   metadata_size, &metadata_flags);
534 	if (flags) {
535 		*flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
536 				KFD_IOC_ALLOC_MEM_FLAGS_VRAM
537 				: KFD_IOC_ALLOC_MEM_FLAGS_GTT;
538 
539 		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
540 			*flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC;
541 	}
542 
543 out_put:
544 	dma_buf_put(dma_buf);
545 	return r;
546 }
547 
548 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst,
549 					  struct amdgpu_device *src)
550 {
551 	struct amdgpu_device *peer_adev = src;
552 	struct amdgpu_device *adev = dst;
553 	int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev);
554 
555 	if (ret < 0) {
556 		DRM_ERROR("amdgpu: failed to get  xgmi hops count between node %d and %d. ret = %d\n",
557 			adev->gmc.xgmi.physical_node_id,
558 			peer_adev->gmc.xgmi.physical_node_id, ret);
559 		ret = 0;
560 	}
561 	return  (uint8_t)ret;
562 }
563 
564 int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst,
565 					    struct amdgpu_device *src,
566 					    bool is_min)
567 {
568 	struct amdgpu_device *adev = dst, *peer_adev;
569 	int num_links;
570 
571 	if (adev->asic_type != CHIP_ALDEBARAN)
572 		return 0;
573 
574 	if (src)
575 		peer_adev = src;
576 
577 	/* num links returns 0 for indirect peers since indirect route is unknown. */
578 	num_links = is_min ? 1 : amdgpu_xgmi_get_num_links(adev, peer_adev);
579 	if (num_links < 0) {
580 		DRM_ERROR("amdgpu: failed to get xgmi num links between node %d and %d. ret = %d\n",
581 			adev->gmc.xgmi.physical_node_id,
582 			peer_adev->gmc.xgmi.physical_node_id, num_links);
583 		num_links = 0;
584 	}
585 
586 	/* Aldebaran xGMI DPM is defeatured so assume x16 x 25Gbps for bandwidth. */
587 	return (num_links * 16 * 25000)/BITS_PER_BYTE;
588 }
589 
590 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min)
591 {
592 	int num_lanes_shift = (is_min ? ffs(adev->pm.pcie_mlw_mask) :
593 							fls(adev->pm.pcie_mlw_mask)) - 1;
594 	int gen_speed_shift = (is_min ? ffs(adev->pm.pcie_gen_mask &
595 						CAIL_PCIE_LINK_SPEED_SUPPORT_MASK) :
596 					fls(adev->pm.pcie_gen_mask &
597 						CAIL_PCIE_LINK_SPEED_SUPPORT_MASK)) - 1;
598 	uint32_t num_lanes_mask = 1 << num_lanes_shift;
599 	uint32_t gen_speed_mask = 1 << gen_speed_shift;
600 	int num_lanes_factor = 0, gen_speed_mbits_factor = 0;
601 
602 	switch (num_lanes_mask) {
603 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X1:
604 		num_lanes_factor = 1;
605 		break;
606 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X2:
607 		num_lanes_factor = 2;
608 		break;
609 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X4:
610 		num_lanes_factor = 4;
611 		break;
612 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X8:
613 		num_lanes_factor = 8;
614 		break;
615 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X12:
616 		num_lanes_factor = 12;
617 		break;
618 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X16:
619 		num_lanes_factor = 16;
620 		break;
621 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X32:
622 		num_lanes_factor = 32;
623 		break;
624 	}
625 
626 	switch (gen_speed_mask) {
627 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1:
628 		gen_speed_mbits_factor = 2500;
629 		break;
630 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2:
631 		gen_speed_mbits_factor = 5000;
632 		break;
633 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3:
634 		gen_speed_mbits_factor = 8000;
635 		break;
636 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4:
637 		gen_speed_mbits_factor = 16000;
638 		break;
639 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5:
640 		gen_speed_mbits_factor = 32000;
641 		break;
642 	}
643 
644 	return (num_lanes_factor * gen_speed_mbits_factor)/BITS_PER_BYTE;
645 }
646 
647 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev,
648 				enum kgd_engine_type engine,
649 				uint32_t vmid, uint64_t gpu_addr,
650 				uint32_t *ib_cmd, uint32_t ib_len)
651 {
652 	struct amdgpu_job *job;
653 	struct amdgpu_ib *ib;
654 	struct amdgpu_ring *ring;
655 	struct dma_fence *f = NULL;
656 	int ret;
657 
658 	switch (engine) {
659 	case KGD_ENGINE_MEC1:
660 		ring = &adev->gfx.compute_ring[0];
661 		break;
662 	case KGD_ENGINE_SDMA1:
663 		ring = &adev->sdma.instance[0].ring;
664 		break;
665 	case KGD_ENGINE_SDMA2:
666 		ring = &adev->sdma.instance[1].ring;
667 		break;
668 	default:
669 		pr_err("Invalid engine in IB submission: %d\n", engine);
670 		ret = -EINVAL;
671 		goto err;
672 	}
673 
674 	ret = amdgpu_job_alloc(adev, 1, &job, NULL);
675 	if (ret)
676 		goto err;
677 
678 	ib = &job->ibs[0];
679 	memset(ib, 0, sizeof(struct amdgpu_ib));
680 
681 	ib->gpu_addr = gpu_addr;
682 	ib->ptr = ib_cmd;
683 	ib->length_dw = ib_len;
684 	/* This works for NO_HWS. TODO: need to handle without knowing VMID */
685 	job->vmid = vmid;
686 
687 	ret = amdgpu_ib_schedule(ring, 1, ib, job, &f);
688 
689 	if (ret) {
690 		DRM_ERROR("amdgpu: failed to schedule IB.\n");
691 		goto err_ib_sched;
692 	}
693 
694 	/* Drop the initial kref_init count (see drm_sched_main as example) */
695 	dma_fence_put(f);
696 	ret = dma_fence_wait(f, false);
697 
698 err_ib_sched:
699 	amdgpu_job_free(job);
700 err:
701 	return ret;
702 }
703 
704 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle)
705 {
706 	amdgpu_dpm_switch_power_profile(adev,
707 					PP_SMC_POWER_PROFILE_COMPUTE,
708 					!idle);
709 }
710 
711 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
712 {
713 	if (adev->kfd.dev)
714 		return vmid >= adev->vm_manager.first_kfd_vmid;
715 
716 	return false;
717 }
718 
719 int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev,
720 				     uint16_t vmid)
721 {
722 	if (adev->family == AMDGPU_FAMILY_AI) {
723 		int i;
724 
725 		for (i = 0; i < adev->num_vmhubs; i++)
726 			amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
727 	} else {
728 		amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0);
729 	}
730 
731 	return 0;
732 }
733 
734 int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
735 				      uint16_t pasid, enum TLB_FLUSH_TYPE flush_type)
736 {
737 	bool all_hub = false;
738 
739 	if (adev->family == AMDGPU_FAMILY_AI ||
740 	    adev->family == AMDGPU_FAMILY_RV)
741 		all_hub = true;
742 
743 	return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub);
744 }
745 
746 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev)
747 {
748 	return adev->have_atomics_support;
749 }
750 
751 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, bool reset)
752 {
753 	struct ras_err_data err_data = {0, 0, 0, NULL};
754 
755 	/* CPU MCA will handle page retirement if connected_to_cpu is 1 */
756 	if (!adev->gmc.xgmi.connected_to_cpu)
757 		amdgpu_umc_poison_handler(adev, &err_data, reset);
758 	else if (reset)
759 		amdgpu_amdkfd_gpu_reset(adev);
760 }
761 
762 bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev)
763 {
764 	if (adev->gfx.ras && adev->gfx.ras->query_utcl2_poison_status)
765 		return adev->gfx.ras->query_utcl2_poison_status(adev);
766 	else
767 		return false;
768 }
769