1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright 2014 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
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 COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include "amdgpu_amdkfd.h"
25 #include "amd_pcie.h"
26 #include "amd_shared.h"
27 
28 #include "amdgpu.h"
29 #include "amdgpu_gfx.h"
30 #include "amdgpu_dma_buf.h"
31 #include <linux/module.h>
32 #include <linux/dma-buf.h>
33 #include "amdgpu_xgmi.h"
34 #include <uapi/linux/kfd_ioctl.h>
35 #include "amdgpu_ras.h"
36 #include "amdgpu_umc.h"
37 #include "amdgpu_reset.h"
38 
39 /* Total memory size in system memory and all GPU VRAM. Used to
40  * estimate worst case amount of memory to reserve for page tables
41  */
42 uint64_t amdgpu_amdkfd_total_mem_size;
43 
44 static bool kfd_initialized;
45 
46 int amdgpu_amdkfd_init(void)
47 {
48 	struct sysinfo si;
49 	int ret;
50 
51 	si_meminfo(&si);
52 	amdgpu_amdkfd_total_mem_size = si.freeram - si.freehigh;
53 	amdgpu_amdkfd_total_mem_size *= si.mem_unit;
54 
55 	ret = kgd2kfd_init();
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 
79 /**
80  * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
81  *                                setup amdkfd
82  *
83  * @adev: amdgpu_device pointer
84  * @aperture_base: output returning doorbell aperture base physical address
85  * @aperture_size: output returning doorbell aperture size in bytes
86  * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
87  *
88  * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
89  * takes doorbells required for its own rings and reports the setup to amdkfd.
90  * amdgpu reserved doorbells are at the start of the doorbell aperture.
91  */
92 static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
93 					 phys_addr_t *aperture_base,
94 					 size_t *aperture_size,
95 					 size_t *start_offset)
96 {
97 	/*
98 	 * The first num_kernel_doorbells are used by amdgpu.
99 	 * amdkfd takes whatever's left in the aperture.
100 	 */
101 	if (adev->enable_mes) {
102 		/*
103 		 * With MES enabled, we only need to initialize
104 		 * the base address. The size and offset are
105 		 * not initialized as AMDGPU manages the whole
106 		 * doorbell space.
107 		 */
108 		*aperture_base = adev->doorbell.base;
109 		*aperture_size = 0;
110 		*start_offset = 0;
111 	} else if (adev->doorbell.size > adev->doorbell.num_kernel_doorbells *
112 						sizeof(u32)) {
113 		*aperture_base = adev->doorbell.base;
114 		*aperture_size = adev->doorbell.size;
115 		*start_offset = adev->doorbell.num_kernel_doorbells * sizeof(u32);
116 	} else {
117 		*aperture_base = 0;
118 		*aperture_size = 0;
119 		*start_offset = 0;
120 	}
121 }
122 
123 
124 static void amdgpu_amdkfd_reset_work(struct work_struct *work)
125 {
126 	struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
127 						  kfd.reset_work);
128 
129 	struct amdgpu_reset_context reset_context;
130 
131 	memset(&reset_context, 0, sizeof(reset_context));
132 
133 	reset_context.method = AMD_RESET_METHOD_NONE;
134 	reset_context.reset_req_dev = adev;
135 	clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
136 
137 	amdgpu_device_gpu_recover(adev, NULL, &reset_context);
138 }
139 
140 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
141 {
142 	int i;
143 	int last_valid_bit;
144 
145 	amdgpu_amdkfd_gpuvm_init_mem_limits();
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_bitmap[0].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 							&gpu_resources);
200 
201 		amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
202 
203 		INIT_WORK(&adev->kfd.reset_work, amdgpu_amdkfd_reset_work);
204 	}
205 }
206 
207 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev)
208 {
209 	if (adev->kfd.dev) {
210 		kgd2kfd_device_exit(adev->kfd.dev);
211 		adev->kfd.dev = NULL;
212 		amdgpu_amdkfd_total_mem_size -= adev->gmc.real_vram_size;
213 	}
214 }
215 
216 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
217 		const void *ih_ring_entry)
218 {
219 	if (adev->kfd.dev)
220 		kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
221 }
222 
223 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm)
224 {
225 	if (adev->kfd.dev)
226 		kgd2kfd_suspend(adev->kfd.dev, run_pm);
227 }
228 
229 int amdgpu_amdkfd_resume_iommu(struct amdgpu_device *adev)
230 {
231 	int r = 0;
232 
233 	if (adev->kfd.dev)
234 		r = kgd2kfd_resume_iommu(adev->kfd.dev);
235 
236 	return r;
237 }
238 
239 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm)
240 {
241 	int r = 0;
242 
243 	if (adev->kfd.dev)
244 		r = kgd2kfd_resume(adev->kfd.dev, run_pm);
245 
246 	return r;
247 }
248 
249 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
250 {
251 	int r = 0;
252 
253 	if (adev->kfd.dev)
254 		r = kgd2kfd_pre_reset(adev->kfd.dev);
255 
256 	return r;
257 }
258 
259 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
260 {
261 	int r = 0;
262 
263 	if (adev->kfd.dev)
264 		r = kgd2kfd_post_reset(adev->kfd.dev);
265 
266 	return r;
267 }
268 
269 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev)
270 {
271 	if (amdgpu_device_should_recover_gpu(adev))
272 		amdgpu_reset_domain_schedule(adev->reset_domain,
273 					     &adev->kfd.reset_work);
274 }
275 
276 int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size,
277 				void **mem_obj, uint64_t *gpu_addr,
278 				void **cpu_ptr, bool cp_mqd_gfx9)
279 {
280 	struct amdgpu_bo *bo = NULL;
281 	struct amdgpu_bo_param bp;
282 	int r;
283 	void *cpu_ptr_tmp = NULL;
284 
285 	memset(&bp, 0, sizeof(bp));
286 	bp.size = size;
287 	bp.byte_align = PAGE_SIZE;
288 	bp.domain = AMDGPU_GEM_DOMAIN_GTT;
289 	bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
290 	bp.type = ttm_bo_type_kernel;
291 	bp.resv = NULL;
292 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
293 
294 	if (cp_mqd_gfx9)
295 		bp.flags |= AMDGPU_GEM_CREATE_CP_MQD_GFX9;
296 
297 	r = amdgpu_bo_create(adev, &bp, &bo);
298 	if (r) {
299 		dev_err(adev->dev,
300 			"failed to allocate BO for amdkfd (%d)\n", r);
301 		return r;
302 	}
303 
304 	/* map the buffer */
305 	r = amdgpu_bo_reserve(bo, true);
306 	if (r) {
307 		dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
308 		goto allocate_mem_reserve_bo_failed;
309 	}
310 
311 	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
312 	if (r) {
313 		dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r);
314 		goto allocate_mem_pin_bo_failed;
315 	}
316 
317 	r = amdgpu_ttm_alloc_gart(&bo->tbo);
318 	if (r) {
319 		dev_err(adev->dev, "%p bind failed\n", bo);
320 		goto allocate_mem_kmap_bo_failed;
321 	}
322 
323 	r = amdgpu_bo_kmap(bo, &cpu_ptr_tmp);
324 	if (r) {
325 		dev_err(adev->dev,
326 			"(%d) failed to map bo to kernel for amdkfd\n", r);
327 		goto allocate_mem_kmap_bo_failed;
328 	}
329 
330 	*mem_obj = bo;
331 	*gpu_addr = amdgpu_bo_gpu_offset(bo);
332 	*cpu_ptr = cpu_ptr_tmp;
333 
334 	amdgpu_bo_unreserve(bo);
335 
336 	return 0;
337 
338 allocate_mem_kmap_bo_failed:
339 	amdgpu_bo_unpin(bo);
340 allocate_mem_pin_bo_failed:
341 	amdgpu_bo_unreserve(bo);
342 allocate_mem_reserve_bo_failed:
343 	amdgpu_bo_unref(&bo);
344 
345 	return r;
346 }
347 
348 void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void *mem_obj)
349 {
350 	struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
351 
352 	amdgpu_bo_reserve(bo, true);
353 	amdgpu_bo_kunmap(bo);
354 	amdgpu_bo_unpin(bo);
355 	amdgpu_bo_unreserve(bo);
356 	amdgpu_bo_unref(&(bo));
357 }
358 
359 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size,
360 				void **mem_obj)
361 {
362 	struct amdgpu_bo *bo = NULL;
363 	struct amdgpu_bo_user *ubo;
364 	struct amdgpu_bo_param bp;
365 	int r;
366 
367 	memset(&bp, 0, sizeof(bp));
368 	bp.size = size;
369 	bp.byte_align = 1;
370 	bp.domain = AMDGPU_GEM_DOMAIN_GWS;
371 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
372 	bp.type = ttm_bo_type_device;
373 	bp.resv = NULL;
374 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
375 
376 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
377 	if (r) {
378 		dev_err(adev->dev,
379 			"failed to allocate gws BO for amdkfd (%d)\n", r);
380 		return r;
381 	}
382 
383 	bo = &ubo->bo;
384 	*mem_obj = bo;
385 	return 0;
386 }
387 
388 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj)
389 {
390 	struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj;
391 
392 	amdgpu_bo_unref(&bo);
393 }
394 
395 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev,
396 				      enum kgd_engine_type type)
397 {
398 	switch (type) {
399 	case KGD_ENGINE_PFP:
400 		return adev->gfx.pfp_fw_version;
401 
402 	case KGD_ENGINE_ME:
403 		return adev->gfx.me_fw_version;
404 
405 	case KGD_ENGINE_CE:
406 		return adev->gfx.ce_fw_version;
407 
408 	case KGD_ENGINE_MEC1:
409 		return adev->gfx.mec_fw_version;
410 
411 	case KGD_ENGINE_MEC2:
412 		return adev->gfx.mec2_fw_version;
413 
414 	case KGD_ENGINE_RLC:
415 		return adev->gfx.rlc_fw_version;
416 
417 	case KGD_ENGINE_SDMA1:
418 		return adev->sdma.instance[0].fw_version;
419 
420 	case KGD_ENGINE_SDMA2:
421 		return adev->sdma.instance[1].fw_version;
422 
423 	default:
424 		return 0;
425 	}
426 
427 	return 0;
428 }
429 
430 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
431 				      struct kfd_local_mem_info *mem_info,
432 				      struct amdgpu_xcp *xcp)
433 {
434 	memset(mem_info, 0, sizeof(*mem_info));
435 
436 	if (xcp) {
437 		if (adev->gmc.real_vram_size == adev->gmc.visible_vram_size)
438 			mem_info->local_mem_size_public =
439 					KFD_XCP_MEMORY_SIZE(adev, xcp->id);
440 		else
441 			mem_info->local_mem_size_private =
442 					KFD_XCP_MEMORY_SIZE(adev, xcp->id);
443 	} else {
444 		mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
445 		mem_info->local_mem_size_private = adev->gmc.real_vram_size -
446 						adev->gmc.visible_vram_size;
447 	}
448 	mem_info->vram_width = adev->gmc.vram_width;
449 
450 	pr_debug("Address base: %pap public 0x%llx private 0x%llx\n",
451 			&adev->gmc.aper_base,
452 			mem_info->local_mem_size_public,
453 			mem_info->local_mem_size_private);
454 
455 	if (amdgpu_sriov_vf(adev))
456 		mem_info->mem_clk_max = adev->clock.default_mclk / 100;
457 	else if (adev->pm.dpm_enabled) {
458 		if (amdgpu_emu_mode == 1)
459 			mem_info->mem_clk_max = 0;
460 		else
461 			mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
462 	} else
463 		mem_info->mem_clk_max = 100;
464 }
465 
466 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev)
467 {
468 	if (adev->gfx.funcs->get_gpu_clock_counter)
469 		return adev->gfx.funcs->get_gpu_clock_counter(adev);
470 	return 0;
471 }
472 
473 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev)
474 {
475 	/* the sclk is in quantas of 10kHz */
476 	if (amdgpu_sriov_vf(adev))
477 		return adev->clock.default_sclk / 100;
478 	else if (adev->pm.dpm_enabled)
479 		return amdgpu_dpm_get_sclk(adev, false) / 100;
480 	else
481 		return 100;
482 }
483 
484 void amdgpu_amdkfd_get_cu_info(struct amdgpu_device *adev, struct kfd_cu_info *cu_info)
485 {
486 	struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
487 
488 	memset(cu_info, 0, sizeof(*cu_info));
489 	if (sizeof(cu_info->cu_bitmap) != sizeof(acu_info.bitmap))
490 		return;
491 
492 	cu_info->cu_active_number = acu_info.number;
493 	cu_info->cu_ao_mask = acu_info.ao_cu_mask;
494 	memcpy(&cu_info->cu_bitmap[0], &acu_info.bitmap[0],
495 	       sizeof(acu_info.bitmap));
496 	cu_info->num_shader_engines = adev->gfx.config.max_shader_engines;
497 	cu_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se;
498 	cu_info->num_cu_per_sh = adev->gfx.config.max_cu_per_sh;
499 	cu_info->simd_per_cu = acu_info.simd_per_cu;
500 	cu_info->max_waves_per_simd = acu_info.max_waves_per_simd;
501 	cu_info->wave_front_size = acu_info.wave_front_size;
502 	cu_info->max_scratch_slots_per_cu = acu_info.max_scratch_slots_per_cu;
503 	cu_info->lds_size = acu_info.lds_size;
504 }
505 
506 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd,
507 				  struct amdgpu_device **dmabuf_adev,
508 				  uint64_t *bo_size, void *metadata_buffer,
509 				  size_t buffer_size, uint32_t *metadata_size,
510 				  uint32_t *flags, int8_t *xcp_id)
511 {
512 	struct dma_buf *dma_buf;
513 	struct drm_gem_object *obj;
514 	struct amdgpu_bo *bo;
515 	uint64_t metadata_flags;
516 	int r = -EINVAL;
517 
518 	dma_buf = dma_buf_get(dma_buf_fd);
519 	if (IS_ERR(dma_buf))
520 		return PTR_ERR(dma_buf);
521 
522 	if (dma_buf->ops != &amdgpu_dmabuf_ops)
523 		/* Can't handle non-graphics buffers */
524 		goto out_put;
525 
526 	obj = dma_buf->priv;
527 	if (obj->dev->driver != adev_to_drm(adev)->driver)
528 		/* Can't handle buffers from different drivers */
529 		goto out_put;
530 
531 	adev = drm_to_adev(obj->dev);
532 	bo = gem_to_amdgpu_bo(obj);
533 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
534 				    AMDGPU_GEM_DOMAIN_GTT)))
535 		/* Only VRAM and GTT BOs are supported */
536 		goto out_put;
537 
538 	r = 0;
539 	if (dmabuf_adev)
540 		*dmabuf_adev = adev;
541 	if (bo_size)
542 		*bo_size = amdgpu_bo_size(bo);
543 	if (metadata_buffer)
544 		r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
545 					   metadata_size, &metadata_flags);
546 	if (flags) {
547 		*flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
548 				KFD_IOC_ALLOC_MEM_FLAGS_VRAM
549 				: KFD_IOC_ALLOC_MEM_FLAGS_GTT;
550 
551 		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
552 			*flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC;
553 	}
554 	if (xcp_id)
555 		*xcp_id = bo->xcp_id;
556 
557 out_put:
558 	dma_buf_put(dma_buf);
559 	return r;
560 }
561 
562 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst,
563 					  struct amdgpu_device *src)
564 {
565 	struct amdgpu_device *peer_adev = src;
566 	struct amdgpu_device *adev = dst;
567 	int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev);
568 
569 	if (ret < 0) {
570 		DRM_ERROR("amdgpu: failed to get  xgmi hops count between node %d and %d. ret = %d\n",
571 			adev->gmc.xgmi.physical_node_id,
572 			peer_adev->gmc.xgmi.physical_node_id, ret);
573 		ret = 0;
574 	}
575 	return  (uint8_t)ret;
576 }
577 
578 int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst,
579 					    struct amdgpu_device *src,
580 					    bool is_min)
581 {
582 	struct amdgpu_device *adev = dst, *peer_adev;
583 	int num_links;
584 
585 	if (adev->asic_type != CHIP_ALDEBARAN)
586 		return 0;
587 
588 	if (src)
589 		peer_adev = src;
590 
591 	/* num links returns 0 for indirect peers since indirect route is unknown. */
592 	num_links = is_min ? 1 : amdgpu_xgmi_get_num_links(adev, peer_adev);
593 	if (num_links < 0) {
594 		DRM_ERROR("amdgpu: failed to get xgmi num links between node %d and %d. ret = %d\n",
595 			adev->gmc.xgmi.physical_node_id,
596 			peer_adev->gmc.xgmi.physical_node_id, num_links);
597 		num_links = 0;
598 	}
599 
600 	/* Aldebaran xGMI DPM is defeatured so assume x16 x 25Gbps for bandwidth. */
601 	return (num_links * 16 * 25000)/BITS_PER_BYTE;
602 }
603 
604 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min)
605 {
606 	int num_lanes_shift = (is_min ? ffs(adev->pm.pcie_mlw_mask) :
607 							fls(adev->pm.pcie_mlw_mask)) - 1;
608 	int gen_speed_shift = (is_min ? ffs(adev->pm.pcie_gen_mask &
609 						CAIL_PCIE_LINK_SPEED_SUPPORT_MASK) :
610 					fls(adev->pm.pcie_gen_mask &
611 						CAIL_PCIE_LINK_SPEED_SUPPORT_MASK)) - 1;
612 	uint32_t num_lanes_mask = 1 << num_lanes_shift;
613 	uint32_t gen_speed_mask = 1 << gen_speed_shift;
614 	int num_lanes_factor = 0, gen_speed_mbits_factor = 0;
615 
616 	switch (num_lanes_mask) {
617 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X1:
618 		num_lanes_factor = 1;
619 		break;
620 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X2:
621 		num_lanes_factor = 2;
622 		break;
623 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X4:
624 		num_lanes_factor = 4;
625 		break;
626 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X8:
627 		num_lanes_factor = 8;
628 		break;
629 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X12:
630 		num_lanes_factor = 12;
631 		break;
632 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X16:
633 		num_lanes_factor = 16;
634 		break;
635 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X32:
636 		num_lanes_factor = 32;
637 		break;
638 	}
639 
640 	switch (gen_speed_mask) {
641 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1:
642 		gen_speed_mbits_factor = 2500;
643 		break;
644 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2:
645 		gen_speed_mbits_factor = 5000;
646 		break;
647 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3:
648 		gen_speed_mbits_factor = 8000;
649 		break;
650 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4:
651 		gen_speed_mbits_factor = 16000;
652 		break;
653 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5:
654 		gen_speed_mbits_factor = 32000;
655 		break;
656 	}
657 
658 	return (num_lanes_factor * gen_speed_mbits_factor)/BITS_PER_BYTE;
659 }
660 
661 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev,
662 				enum kgd_engine_type engine,
663 				uint32_t vmid, uint64_t gpu_addr,
664 				uint32_t *ib_cmd, uint32_t ib_len)
665 {
666 	struct amdgpu_job *job;
667 	struct amdgpu_ib *ib;
668 	struct amdgpu_ring *ring;
669 	struct dma_fence *f = NULL;
670 	int ret;
671 
672 	switch (engine) {
673 	case KGD_ENGINE_MEC1:
674 		ring = &adev->gfx.compute_ring[0];
675 		break;
676 	case KGD_ENGINE_SDMA1:
677 		ring = &adev->sdma.instance[0].ring;
678 		break;
679 	case KGD_ENGINE_SDMA2:
680 		ring = &adev->sdma.instance[1].ring;
681 		break;
682 	default:
683 		pr_err("Invalid engine in IB submission: %d\n", engine);
684 		ret = -EINVAL;
685 		goto err;
686 	}
687 
688 	ret = amdgpu_job_alloc(adev, NULL, NULL, NULL, 1, &job);
689 	if (ret)
690 		goto err;
691 
692 	ib = &job->ibs[0];
693 	memset(ib, 0, sizeof(struct amdgpu_ib));
694 
695 	ib->gpu_addr = gpu_addr;
696 	ib->ptr = ib_cmd;
697 	ib->length_dw = ib_len;
698 	/* This works for NO_HWS. TODO: need to handle without knowing VMID */
699 	job->vmid = vmid;
700 	job->num_ibs = 1;
701 
702 	ret = amdgpu_ib_schedule(ring, 1, ib, job, &f);
703 
704 	if (ret) {
705 		DRM_ERROR("amdgpu: failed to schedule IB.\n");
706 		goto err_ib_sched;
707 	}
708 
709 	/* Drop the initial kref_init count (see drm_sched_main as example) */
710 	dma_fence_put(f);
711 	ret = dma_fence_wait(f, false);
712 
713 err_ib_sched:
714 	amdgpu_job_free(job);
715 err:
716 	return ret;
717 }
718 
719 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle)
720 {
721 	/* Temporary workaround to fix issues observed in some
722 	 * compute applications when GFXOFF is enabled on GFX11.
723 	 */
724 	if (IP_VERSION_MAJ(adev->ip_versions[GC_HWIP][0]) == 11) {
725 		pr_debug("GFXOFF is %s\n", idle ? "enabled" : "disabled");
726 		amdgpu_gfx_off_ctrl(adev, idle);
727 	}
728 	amdgpu_dpm_switch_power_profile(adev,
729 					PP_SMC_POWER_PROFILE_COMPUTE,
730 					!idle);
731 }
732 
733 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
734 {
735 	if (adev->kfd.dev)
736 		return vmid >= adev->vm_manager.first_kfd_vmid;
737 
738 	return false;
739 }
740 
741 int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct amdgpu_device *adev,
742 				     uint16_t vmid)
743 {
744 	if (adev->family == AMDGPU_FAMILY_AI) {
745 		int i;
746 
747 		for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS)
748 			amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
749 	} else {
750 		amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB(0), 0);
751 	}
752 
753 	return 0;
754 }
755 
756 int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
757 				      uint16_t pasid,
758 				      enum TLB_FLUSH_TYPE flush_type,
759 				      uint32_t inst)
760 {
761 	bool all_hub = false;
762 
763 	if (adev->family == AMDGPU_FAMILY_AI ||
764 	    adev->family == AMDGPU_FAMILY_RV)
765 		all_hub = true;
766 
767 	return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub, inst);
768 }
769 
770 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev)
771 {
772 	return adev->have_atomics_support;
773 }
774 
775 void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev)
776 {
777 	amdgpu_device_flush_hdp(adev, NULL);
778 }
779 
780 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev, bool reset)
781 {
782 	amdgpu_umc_poison_handler(adev, reset);
783 }
784 
785 int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev,
786 					uint32_t *payload)
787 {
788 	int ret;
789 
790 	/* Device or IH ring is not ready so bail. */
791 	ret = amdgpu_ih_wait_on_checkpoint_process_ts(adev, &adev->irq.ih);
792 	if (ret)
793 		return ret;
794 
795 	/* Send payload to fence KFD interrupts */
796 	amdgpu_amdkfd_interrupt(adev, payload);
797 
798 	return 0;
799 }
800 
801 bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev)
802 {
803 	if (adev->gfx.ras && adev->gfx.ras->query_utcl2_poison_status)
804 		return adev->gfx.ras->query_utcl2_poison_status(adev);
805 	else
806 		return false;
807 }
808 
809 int amdgpu_amdkfd_check_and_lock_kfd(struct amdgpu_device *adev)
810 {
811 	return kgd2kfd_check_and_lock_kfd();
812 }
813 
814 void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev)
815 {
816 	kgd2kfd_unlock_kfd();
817 }
818 
819 
820 u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id)
821 {
822 	u64 tmp;
823 	s8 mem_id = KFD_XCP_MEM_ID(adev, xcp_id);
824 
825 	if (adev->gmc.num_mem_partitions && xcp_id >= 0 && mem_id >= 0) {
826 		tmp = adev->gmc.mem_partitions[mem_id].size;
827 		do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition);
828 		return ALIGN_DOWN(tmp, PAGE_SIZE);
829 	} else {
830 		return adev->gmc.real_vram_size;
831 	}
832 }
833