xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c (revision a23e1966932464e1c5226cb9ac4ce1d5fc10ba22)
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 <drm/ttm/ttm_tt.h>
32 #include <linux/module.h>
33 #include <linux/dma-buf.h>
34 #include "amdgpu_xgmi.h"
35 #include <uapi/linux/kfd_ioctl.h>
36 #include "amdgpu_ras.h"
37 #include "amdgpu_umc.h"
38 #include "amdgpu_reset.h"
39 
40 /* Total memory size in system memory and all GPU VRAM. Used to
41  * estimate worst case amount of memory to reserve for page tables
42  */
43 uint64_t amdgpu_amdkfd_total_mem_size;
44 
45 static bool kfd_initialized;
46 
47 int amdgpu_amdkfd_init(void)
48 {
49 	struct sysinfo si;
50 	int ret;
51 
52 	si_meminfo(&si);
53 	amdgpu_amdkfd_total_mem_size = si.freeram - si.freehigh;
54 	amdgpu_amdkfd_total_mem_size *= si.mem_unit;
55 
56 	ret = kgd2kfd_init();
57 	kfd_initialized = !ret;
58 
59 	return ret;
60 }
61 
62 void amdgpu_amdkfd_fini(void)
63 {
64 	if (kfd_initialized) {
65 		kgd2kfd_exit();
66 		kfd_initialized = false;
67 	}
68 }
69 
70 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
71 {
72 	bool vf = amdgpu_sriov_vf(adev);
73 
74 	if (!kfd_initialized)
75 		return;
76 
77 	adev->kfd.dev = kgd2kfd_probe(adev, vf);
78 }
79 
80 /**
81  * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
82  *                                setup amdkfd
83  *
84  * @adev: amdgpu_device pointer
85  * @aperture_base: output returning doorbell aperture base physical address
86  * @aperture_size: output returning doorbell aperture size in bytes
87  * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
88  *
89  * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
90  * takes doorbells required for its own rings and reports the setup to amdkfd.
91  * amdgpu reserved doorbells are at the start of the doorbell aperture.
92  */
93 static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
94 					 phys_addr_t *aperture_base,
95 					 size_t *aperture_size,
96 					 size_t *start_offset)
97 {
98 	/*
99 	 * The first num_kernel_doorbells are used by amdgpu.
100 	 * amdkfd takes whatever's left in the aperture.
101 	 */
102 	if (adev->enable_mes) {
103 		/*
104 		 * With MES enabled, we only need to initialize
105 		 * the base address. The size and offset are
106 		 * not initialized as AMDGPU manages the whole
107 		 * doorbell space.
108 		 */
109 		*aperture_base = adev->doorbell.base;
110 		*aperture_size = 0;
111 		*start_offset = 0;
112 	} else if (adev->doorbell.size > adev->doorbell.num_kernel_doorbells *
113 						sizeof(u32)) {
114 		*aperture_base = adev->doorbell.base;
115 		*aperture_size = adev->doorbell.size;
116 		*start_offset = adev->doorbell.num_kernel_doorbells * sizeof(u32);
117 	} else {
118 		*aperture_base = 0;
119 		*aperture_size = 0;
120 		*start_offset = 0;
121 	}
122 }
123 
124 
125 static void amdgpu_amdkfd_reset_work(struct work_struct *work)
126 {
127 	struct amdgpu_device *adev = container_of(work, struct amdgpu_device,
128 						  kfd.reset_work);
129 
130 	struct amdgpu_reset_context reset_context;
131 
132 	memset(&reset_context, 0, sizeof(reset_context));
133 
134 	reset_context.method = AMD_RESET_METHOD_NONE;
135 	reset_context.reset_req_dev = adev;
136 	clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
137 
138 	amdgpu_device_gpu_recover(adev, NULL, &reset_context);
139 }
140 
141 static const struct drm_client_funcs kfd_client_funcs = {
142 	.unregister	= drm_client_release,
143 };
144 
145 int amdgpu_amdkfd_drm_client_create(struct amdgpu_device *adev)
146 {
147 	int ret;
148 
149 	if (!adev->kfd.init_complete || adev->kfd.client.dev)
150 		return 0;
151 
152 	ret = drm_client_init(&adev->ddev, &adev->kfd.client, "kfd",
153 			      &kfd_client_funcs);
154 	if (ret) {
155 		dev_err(adev->dev, "Failed to init DRM client: %d\n",
156 			ret);
157 		return ret;
158 	}
159 
160 	drm_client_register(&adev->kfd.client);
161 
162 	return 0;
163 }
164 
165 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
166 {
167 	int i;
168 	int last_valid_bit;
169 
170 	amdgpu_amdkfd_gpuvm_init_mem_limits();
171 
172 	if (adev->kfd.dev) {
173 		struct kgd2kfd_shared_resources gpu_resources = {
174 			.compute_vmid_bitmap =
175 				((1 << AMDGPU_NUM_VMID) - 1) -
176 				((1 << adev->vm_manager.first_kfd_vmid) - 1),
177 			.num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
178 			.num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe,
179 			.gpuvm_size = min(adev->vm_manager.max_pfn
180 					  << AMDGPU_GPU_PAGE_SHIFT,
181 					  AMDGPU_GMC_HOLE_START),
182 			.drm_render_minor = adev_to_drm(adev)->render->index,
183 			.sdma_doorbell_idx = adev->doorbell_index.sdma_engine,
184 			.enable_mes = adev->enable_mes,
185 		};
186 
187 		/* this is going to have a few of the MSBs set that we need to
188 		 * clear
189 		 */
190 		bitmap_complement(gpu_resources.cp_queue_bitmap,
191 				  adev->gfx.mec_bitmap[0].queue_bitmap,
192 				  AMDGPU_MAX_QUEUES);
193 
194 		/* According to linux/bitmap.h we shouldn't use bitmap_clear if
195 		 * nbits is not compile time constant
196 		 */
197 		last_valid_bit = 1 /* only first MEC can have compute queues */
198 				* adev->gfx.mec.num_pipe_per_mec
199 				* adev->gfx.mec.num_queue_per_pipe;
200 		for (i = last_valid_bit; i < AMDGPU_MAX_QUEUES; ++i)
201 			clear_bit(i, gpu_resources.cp_queue_bitmap);
202 
203 		amdgpu_doorbell_get_kfd_info(adev,
204 				&gpu_resources.doorbell_physical_address,
205 				&gpu_resources.doorbell_aperture_size,
206 				&gpu_resources.doorbell_start_offset);
207 
208 		/* Since SOC15, BIF starts to statically use the
209 		 * lower 12 bits of doorbell addresses for routing
210 		 * based on settings in registers like
211 		 * SDMA0_DOORBELL_RANGE etc..
212 		 * In order to route a doorbell to CP engine, the lower
213 		 * 12 bits of its address has to be outside the range
214 		 * set for SDMA, VCN, and IH blocks.
215 		 */
216 		if (adev->asic_type >= CHIP_VEGA10) {
217 			gpu_resources.non_cp_doorbells_start =
218 					adev->doorbell_index.first_non_cp;
219 			gpu_resources.non_cp_doorbells_end =
220 					adev->doorbell_index.last_non_cp;
221 		}
222 
223 		adev->kfd.init_complete = kgd2kfd_device_init(adev->kfd.dev,
224 							&gpu_resources);
225 
226 		amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
227 
228 		INIT_WORK(&adev->kfd.reset_work, amdgpu_amdkfd_reset_work);
229 	}
230 }
231 
232 void amdgpu_amdkfd_device_fini_sw(struct amdgpu_device *adev)
233 {
234 	if (adev->kfd.dev) {
235 		kgd2kfd_device_exit(adev->kfd.dev);
236 		adev->kfd.dev = NULL;
237 		amdgpu_amdkfd_total_mem_size -= adev->gmc.real_vram_size;
238 	}
239 }
240 
241 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
242 		const void *ih_ring_entry)
243 {
244 	if (adev->kfd.dev)
245 		kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
246 }
247 
248 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm)
249 {
250 	if (adev->kfd.dev)
251 		kgd2kfd_suspend(adev->kfd.dev, run_pm);
252 }
253 
254 int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm)
255 {
256 	int r = 0;
257 
258 	if (adev->kfd.dev)
259 		r = kgd2kfd_resume(adev->kfd.dev, run_pm);
260 
261 	return r;
262 }
263 
264 int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
265 {
266 	int r = 0;
267 
268 	if (adev->kfd.dev)
269 		r = kgd2kfd_pre_reset(adev->kfd.dev);
270 
271 	return r;
272 }
273 
274 int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
275 {
276 	int r = 0;
277 
278 	if (adev->kfd.dev)
279 		r = kgd2kfd_post_reset(adev->kfd.dev);
280 
281 	return r;
282 }
283 
284 void amdgpu_amdkfd_gpu_reset(struct amdgpu_device *adev)
285 {
286 	if (amdgpu_device_should_recover_gpu(adev))
287 		amdgpu_reset_domain_schedule(adev->reset_domain,
288 					     &adev->kfd.reset_work);
289 }
290 
291 int amdgpu_amdkfd_alloc_gtt_mem(struct amdgpu_device *adev, size_t size,
292 				void **mem_obj, uint64_t *gpu_addr,
293 				void **cpu_ptr, bool cp_mqd_gfx9)
294 {
295 	struct amdgpu_bo *bo = NULL;
296 	struct amdgpu_bo_param bp;
297 	int r;
298 	void *cpu_ptr_tmp = NULL;
299 
300 	memset(&bp, 0, sizeof(bp));
301 	bp.size = size;
302 	bp.byte_align = PAGE_SIZE;
303 	bp.domain = AMDGPU_GEM_DOMAIN_GTT;
304 	bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
305 	bp.type = ttm_bo_type_kernel;
306 	bp.resv = NULL;
307 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
308 
309 	if (cp_mqd_gfx9)
310 		bp.flags |= AMDGPU_GEM_CREATE_CP_MQD_GFX9;
311 
312 	r = amdgpu_bo_create(adev, &bp, &bo);
313 	if (r) {
314 		dev_err(adev->dev,
315 			"failed to allocate BO for amdkfd (%d)\n", r);
316 		return r;
317 	}
318 
319 	/* map the buffer */
320 	r = amdgpu_bo_reserve(bo, true);
321 	if (r) {
322 		dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
323 		goto allocate_mem_reserve_bo_failed;
324 	}
325 
326 	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
327 	if (r) {
328 		dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r);
329 		goto allocate_mem_pin_bo_failed;
330 	}
331 
332 	r = amdgpu_ttm_alloc_gart(&bo->tbo);
333 	if (r) {
334 		dev_err(adev->dev, "%p bind failed\n", bo);
335 		goto allocate_mem_kmap_bo_failed;
336 	}
337 
338 	r = amdgpu_bo_kmap(bo, &cpu_ptr_tmp);
339 	if (r) {
340 		dev_err(adev->dev,
341 			"(%d) failed to map bo to kernel for amdkfd\n", r);
342 		goto allocate_mem_kmap_bo_failed;
343 	}
344 
345 	*mem_obj = bo;
346 	*gpu_addr = amdgpu_bo_gpu_offset(bo);
347 	*cpu_ptr = cpu_ptr_tmp;
348 
349 	amdgpu_bo_unreserve(bo);
350 
351 	return 0;
352 
353 allocate_mem_kmap_bo_failed:
354 	amdgpu_bo_unpin(bo);
355 allocate_mem_pin_bo_failed:
356 	amdgpu_bo_unreserve(bo);
357 allocate_mem_reserve_bo_failed:
358 	amdgpu_bo_unref(&bo);
359 
360 	return r;
361 }
362 
363 void amdgpu_amdkfd_free_gtt_mem(struct amdgpu_device *adev, void *mem_obj)
364 {
365 	struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
366 
367 	amdgpu_bo_reserve(bo, true);
368 	amdgpu_bo_kunmap(bo);
369 	amdgpu_bo_unpin(bo);
370 	amdgpu_bo_unreserve(bo);
371 	amdgpu_bo_unref(&(bo));
372 }
373 
374 int amdgpu_amdkfd_alloc_gws(struct amdgpu_device *adev, size_t size,
375 				void **mem_obj)
376 {
377 	struct amdgpu_bo *bo = NULL;
378 	struct amdgpu_bo_user *ubo;
379 	struct amdgpu_bo_param bp;
380 	int r;
381 
382 	memset(&bp, 0, sizeof(bp));
383 	bp.size = size;
384 	bp.byte_align = 1;
385 	bp.domain = AMDGPU_GEM_DOMAIN_GWS;
386 	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
387 	bp.type = ttm_bo_type_device;
388 	bp.resv = NULL;
389 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
390 
391 	r = amdgpu_bo_create_user(adev, &bp, &ubo);
392 	if (r) {
393 		dev_err(adev->dev,
394 			"failed to allocate gws BO for amdkfd (%d)\n", r);
395 		return r;
396 	}
397 
398 	bo = &ubo->bo;
399 	*mem_obj = bo;
400 	return 0;
401 }
402 
403 void amdgpu_amdkfd_free_gws(struct amdgpu_device *adev, void *mem_obj)
404 {
405 	struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj;
406 
407 	amdgpu_bo_unref(&bo);
408 }
409 
410 uint32_t amdgpu_amdkfd_get_fw_version(struct amdgpu_device *adev,
411 				      enum kgd_engine_type type)
412 {
413 	switch (type) {
414 	case KGD_ENGINE_PFP:
415 		return adev->gfx.pfp_fw_version;
416 
417 	case KGD_ENGINE_ME:
418 		return adev->gfx.me_fw_version;
419 
420 	case KGD_ENGINE_CE:
421 		return adev->gfx.ce_fw_version;
422 
423 	case KGD_ENGINE_MEC1:
424 		return adev->gfx.mec_fw_version;
425 
426 	case KGD_ENGINE_MEC2:
427 		return adev->gfx.mec2_fw_version;
428 
429 	case KGD_ENGINE_RLC:
430 		return adev->gfx.rlc_fw_version;
431 
432 	case KGD_ENGINE_SDMA1:
433 		return adev->sdma.instance[0].fw_version;
434 
435 	case KGD_ENGINE_SDMA2:
436 		return adev->sdma.instance[1].fw_version;
437 
438 	default:
439 		return 0;
440 	}
441 
442 	return 0;
443 }
444 
445 void amdgpu_amdkfd_get_local_mem_info(struct amdgpu_device *adev,
446 				      struct kfd_local_mem_info *mem_info,
447 				      struct amdgpu_xcp *xcp)
448 {
449 	memset(mem_info, 0, sizeof(*mem_info));
450 
451 	if (xcp) {
452 		if (adev->gmc.real_vram_size == adev->gmc.visible_vram_size)
453 			mem_info->local_mem_size_public =
454 					KFD_XCP_MEMORY_SIZE(adev, xcp->id);
455 		else
456 			mem_info->local_mem_size_private =
457 					KFD_XCP_MEMORY_SIZE(adev, xcp->id);
458 	} else {
459 		mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
460 		mem_info->local_mem_size_private = adev->gmc.real_vram_size -
461 						adev->gmc.visible_vram_size;
462 	}
463 	mem_info->vram_width = adev->gmc.vram_width;
464 
465 	pr_debug("Address base: %pap public 0x%llx private 0x%llx\n",
466 			&adev->gmc.aper_base,
467 			mem_info->local_mem_size_public,
468 			mem_info->local_mem_size_private);
469 
470 	if (adev->pm.dpm_enabled) {
471 		if (amdgpu_emu_mode == 1)
472 			mem_info->mem_clk_max = 0;
473 		else
474 			mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
475 	} else
476 		mem_info->mem_clk_max = 100;
477 }
478 
479 uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct amdgpu_device *adev)
480 {
481 	if (adev->gfx.funcs->get_gpu_clock_counter)
482 		return adev->gfx.funcs->get_gpu_clock_counter(adev);
483 	return 0;
484 }
485 
486 uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct amdgpu_device *adev)
487 {
488 	/* the sclk is in quantas of 10kHz */
489 	if (adev->pm.dpm_enabled)
490 		return amdgpu_dpm_get_sclk(adev, false) / 100;
491 	else
492 		return 100;
493 }
494 
495 int amdgpu_amdkfd_get_dmabuf_info(struct amdgpu_device *adev, int dma_buf_fd,
496 				  struct amdgpu_device **dmabuf_adev,
497 				  uint64_t *bo_size, void *metadata_buffer,
498 				  size_t buffer_size, uint32_t *metadata_size,
499 				  uint32_t *flags, int8_t *xcp_id)
500 {
501 	struct dma_buf *dma_buf;
502 	struct drm_gem_object *obj;
503 	struct amdgpu_bo *bo;
504 	uint64_t metadata_flags;
505 	int r = -EINVAL;
506 
507 	dma_buf = dma_buf_get(dma_buf_fd);
508 	if (IS_ERR(dma_buf))
509 		return PTR_ERR(dma_buf);
510 
511 	if (dma_buf->ops != &amdgpu_dmabuf_ops)
512 		/* Can't handle non-graphics buffers */
513 		goto out_put;
514 
515 	obj = dma_buf->priv;
516 	if (obj->dev->driver != adev_to_drm(adev)->driver)
517 		/* Can't handle buffers from different drivers */
518 		goto out_put;
519 
520 	adev = drm_to_adev(obj->dev);
521 	bo = gem_to_amdgpu_bo(obj);
522 	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
523 				    AMDGPU_GEM_DOMAIN_GTT)))
524 		/* Only VRAM and GTT BOs are supported */
525 		goto out_put;
526 
527 	r = 0;
528 	if (dmabuf_adev)
529 		*dmabuf_adev = adev;
530 	if (bo_size)
531 		*bo_size = amdgpu_bo_size(bo);
532 	if (metadata_buffer)
533 		r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
534 					   metadata_size, &metadata_flags);
535 	if (flags) {
536 		*flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
537 				KFD_IOC_ALLOC_MEM_FLAGS_VRAM
538 				: KFD_IOC_ALLOC_MEM_FLAGS_GTT;
539 
540 		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
541 			*flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC;
542 	}
543 	if (xcp_id)
544 		*xcp_id = bo->xcp_id;
545 
546 out_put:
547 	dma_buf_put(dma_buf);
548 	return r;
549 }
550 
551 uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct amdgpu_device *dst,
552 					  struct amdgpu_device *src)
553 {
554 	struct amdgpu_device *peer_adev = src;
555 	struct amdgpu_device *adev = dst;
556 	int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev);
557 
558 	if (ret < 0) {
559 		DRM_ERROR("amdgpu: failed to get  xgmi hops count between node %d and %d. ret = %d\n",
560 			adev->gmc.xgmi.physical_node_id,
561 			peer_adev->gmc.xgmi.physical_node_id, ret);
562 		ret = 0;
563 	}
564 	return  (uint8_t)ret;
565 }
566 
567 int amdgpu_amdkfd_get_xgmi_bandwidth_mbytes(struct amdgpu_device *dst,
568 					    struct amdgpu_device *src,
569 					    bool is_min)
570 {
571 	struct amdgpu_device *adev = dst, *peer_adev;
572 	int num_links;
573 
574 	if (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(9, 4, 2))
575 		return 0;
576 
577 	if (src)
578 		peer_adev = src;
579 
580 	/* num links returns 0 for indirect peers since indirect route is unknown. */
581 	num_links = is_min ? 1 : amdgpu_xgmi_get_num_links(adev, peer_adev);
582 	if (num_links < 0) {
583 		DRM_ERROR("amdgpu: failed to get xgmi num links between node %d and %d. ret = %d\n",
584 			adev->gmc.xgmi.physical_node_id,
585 			peer_adev->gmc.xgmi.physical_node_id, num_links);
586 		num_links = 0;
587 	}
588 
589 	/* Aldebaran xGMI DPM is defeatured so assume x16 x 25Gbps for bandwidth. */
590 	return (num_links * 16 * 25000)/BITS_PER_BYTE;
591 }
592 
593 int amdgpu_amdkfd_get_pcie_bandwidth_mbytes(struct amdgpu_device *adev, bool is_min)
594 {
595 	int num_lanes_shift = (is_min ? ffs(adev->pm.pcie_mlw_mask) :
596 							fls(adev->pm.pcie_mlw_mask)) - 1;
597 	int gen_speed_shift = (is_min ? ffs(adev->pm.pcie_gen_mask &
598 						CAIL_PCIE_LINK_SPEED_SUPPORT_MASK) :
599 					fls(adev->pm.pcie_gen_mask &
600 						CAIL_PCIE_LINK_SPEED_SUPPORT_MASK)) - 1;
601 	uint32_t num_lanes_mask = 1 << num_lanes_shift;
602 	uint32_t gen_speed_mask = 1 << gen_speed_shift;
603 	int num_lanes_factor = 0, gen_speed_mbits_factor = 0;
604 
605 	switch (num_lanes_mask) {
606 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X1:
607 		num_lanes_factor = 1;
608 		break;
609 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X2:
610 		num_lanes_factor = 2;
611 		break;
612 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X4:
613 		num_lanes_factor = 4;
614 		break;
615 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X8:
616 		num_lanes_factor = 8;
617 		break;
618 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X12:
619 		num_lanes_factor = 12;
620 		break;
621 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X16:
622 		num_lanes_factor = 16;
623 		break;
624 	case CAIL_PCIE_LINK_WIDTH_SUPPORT_X32:
625 		num_lanes_factor = 32;
626 		break;
627 	}
628 
629 	switch (gen_speed_mask) {
630 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1:
631 		gen_speed_mbits_factor = 2500;
632 		break;
633 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2:
634 		gen_speed_mbits_factor = 5000;
635 		break;
636 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3:
637 		gen_speed_mbits_factor = 8000;
638 		break;
639 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4:
640 		gen_speed_mbits_factor = 16000;
641 		break;
642 	case CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5:
643 		gen_speed_mbits_factor = 32000;
644 		break;
645 	}
646 
647 	return (num_lanes_factor * gen_speed_mbits_factor)/BITS_PER_BYTE;
648 }
649 
650 int amdgpu_amdkfd_submit_ib(struct amdgpu_device *adev,
651 				enum kgd_engine_type engine,
652 				uint32_t vmid, uint64_t gpu_addr,
653 				uint32_t *ib_cmd, uint32_t ib_len)
654 {
655 	struct amdgpu_job *job;
656 	struct amdgpu_ib *ib;
657 	struct amdgpu_ring *ring;
658 	struct dma_fence *f = NULL;
659 	int ret;
660 
661 	switch (engine) {
662 	case KGD_ENGINE_MEC1:
663 		ring = &adev->gfx.compute_ring[0];
664 		break;
665 	case KGD_ENGINE_SDMA1:
666 		ring = &adev->sdma.instance[0].ring;
667 		break;
668 	case KGD_ENGINE_SDMA2:
669 		ring = &adev->sdma.instance[1].ring;
670 		break;
671 	default:
672 		pr_err("Invalid engine in IB submission: %d\n", engine);
673 		ret = -EINVAL;
674 		goto err;
675 	}
676 
677 	ret = amdgpu_job_alloc(adev, NULL, NULL, NULL, 1, &job);
678 	if (ret)
679 		goto err;
680 
681 	ib = &job->ibs[0];
682 	memset(ib, 0, sizeof(struct amdgpu_ib));
683 
684 	ib->gpu_addr = gpu_addr;
685 	ib->ptr = ib_cmd;
686 	ib->length_dw = ib_len;
687 	/* This works for NO_HWS. TODO: need to handle without knowing VMID */
688 	job->vmid = vmid;
689 	job->num_ibs = 1;
690 
691 	ret = amdgpu_ib_schedule(ring, 1, ib, job, &f);
692 
693 	if (ret) {
694 		DRM_ERROR("amdgpu: failed to schedule IB.\n");
695 		goto err_ib_sched;
696 	}
697 
698 	/* Drop the initial kref_init count (see drm_sched_main as example) */
699 	dma_fence_put(f);
700 	ret = dma_fence_wait(f, false);
701 
702 err_ib_sched:
703 	amdgpu_job_free(job);
704 err:
705 	return ret;
706 }
707 
708 void amdgpu_amdkfd_set_compute_idle(struct amdgpu_device *adev, bool idle)
709 {
710 	enum amd_powergating_state state = idle ? AMD_PG_STATE_GATE : AMD_PG_STATE_UNGATE;
711 	if (IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0)) == 11 &&
712 	    ((adev->mes.kiq_version & AMDGPU_MES_VERSION_MASK) <= 64)) {
713 		pr_debug("GFXOFF is %s\n", idle ? "enabled" : "disabled");
714 		amdgpu_gfx_off_ctrl(adev, idle);
715 	} else if ((IP_VERSION_MAJ(amdgpu_ip_version(adev, GC_HWIP, 0)) == 9) &&
716 		(adev->flags & AMD_IS_APU)) {
717 		/* Disable GFXOFF and PG. Temporary workaround
718 		 * to fix some compute applications issue on GFX9.
719 		 */
720 		adev->ip_blocks[AMD_IP_BLOCK_TYPE_GFX].version->funcs->set_powergating_state((void *)adev, state);
721 	}
722 	amdgpu_dpm_switch_power_profile(adev,
723 					PP_SMC_POWER_PROFILE_COMPUTE,
724 					!idle);
725 }
726 
727 bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
728 {
729 	if (adev->kfd.dev)
730 		return vmid >= adev->vm_manager.first_kfd_vmid;
731 
732 	return false;
733 }
734 
735 bool amdgpu_amdkfd_have_atomics_support(struct amdgpu_device *adev)
736 {
737 	return adev->have_atomics_support;
738 }
739 
740 void amdgpu_amdkfd_debug_mem_fence(struct amdgpu_device *adev)
741 {
742 	amdgpu_device_flush_hdp(adev, NULL);
743 }
744 
745 bool amdgpu_amdkfd_is_fed(struct amdgpu_device *adev)
746 {
747 	return amdgpu_ras_get_fed_status(adev);
748 }
749 
750 void amdgpu_amdkfd_ras_poison_consumption_handler(struct amdgpu_device *adev,
751 	enum amdgpu_ras_block block, bool reset)
752 {
753 	amdgpu_umc_poison_handler(adev, block, reset);
754 }
755 
756 int amdgpu_amdkfd_send_close_event_drain_irq(struct amdgpu_device *adev,
757 					uint32_t *payload)
758 {
759 	int ret;
760 
761 	/* Device or IH ring is not ready so bail. */
762 	ret = amdgpu_ih_wait_on_checkpoint_process_ts(adev, &adev->irq.ih);
763 	if (ret)
764 		return ret;
765 
766 	/* Send payload to fence KFD interrupts */
767 	amdgpu_amdkfd_interrupt(adev, payload);
768 
769 	return 0;
770 }
771 
772 bool amdgpu_amdkfd_ras_query_utcl2_poison_status(struct amdgpu_device *adev)
773 {
774 	if (adev->gfx.ras && adev->gfx.ras->query_utcl2_poison_status)
775 		return adev->gfx.ras->query_utcl2_poison_status(adev);
776 	else
777 		return false;
778 }
779 
780 int amdgpu_amdkfd_check_and_lock_kfd(struct amdgpu_device *adev)
781 {
782 	return kgd2kfd_check_and_lock_kfd();
783 }
784 
785 void amdgpu_amdkfd_unlock_kfd(struct amdgpu_device *adev)
786 {
787 	kgd2kfd_unlock_kfd();
788 }
789 
790 
791 u64 amdgpu_amdkfd_xcp_memory_size(struct amdgpu_device *adev, int xcp_id)
792 {
793 	s8 mem_id = KFD_XCP_MEM_ID(adev, xcp_id);
794 	u64 tmp;
795 
796 	if (adev->gmc.num_mem_partitions && xcp_id >= 0 && mem_id >= 0) {
797 		if (adev->gmc.is_app_apu && adev->gmc.num_mem_partitions == 1) {
798 			/* In NPS1 mode, we should restrict the vram reporting
799 			 * tied to the ttm_pages_limit which is 1/2 of the system
800 			 * memory. For other partition modes, the HBM is uniformly
801 			 * divided already per numa node reported. If user wants to
802 			 * go beyond the default ttm limit and maximize the ROCm
803 			 * allocations, they can go up to max ttm and sysmem limits.
804 			 */
805 
806 			tmp = (ttm_tt_pages_limit() << PAGE_SHIFT) / num_online_nodes();
807 		} else {
808 			tmp = adev->gmc.mem_partitions[mem_id].size;
809 		}
810 		do_div(tmp, adev->xcp_mgr->num_xcp_per_mem_partition);
811 		return ALIGN_DOWN(tmp, PAGE_SIZE);
812 	} else {
813 		return adev->gmc.real_vram_size;
814 	}
815 }
816 
817 int amdgpu_amdkfd_unmap_hiq(struct amdgpu_device *adev, u32 doorbell_off,
818 			    u32 inst)
819 {
820 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst];
821 	struct amdgpu_ring *kiq_ring = &kiq->ring;
822 	struct amdgpu_ring_funcs *ring_funcs;
823 	struct amdgpu_ring *ring;
824 	int r = 0;
825 
826 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
827 		return -EINVAL;
828 
829 	ring_funcs = kzalloc(sizeof(*ring_funcs), GFP_KERNEL);
830 	if (!ring_funcs)
831 		return -ENOMEM;
832 
833 	ring = kzalloc(sizeof(*ring), GFP_KERNEL);
834 	if (!ring) {
835 		r = -ENOMEM;
836 		goto free_ring_funcs;
837 	}
838 
839 	ring_funcs->type = AMDGPU_RING_TYPE_COMPUTE;
840 	ring->doorbell_index = doorbell_off;
841 	ring->funcs = ring_funcs;
842 
843 	spin_lock(&kiq->ring_lock);
844 
845 	if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size)) {
846 		spin_unlock(&kiq->ring_lock);
847 		r = -ENOMEM;
848 		goto free_ring;
849 	}
850 
851 	kiq->pmf->kiq_unmap_queues(kiq_ring, ring, RESET_QUEUES, 0, 0);
852 
853 	if (kiq_ring->sched.ready && !adev->job_hang)
854 		r = amdgpu_ring_test_helper(kiq_ring);
855 
856 	spin_unlock(&kiq->ring_lock);
857 
858 free_ring:
859 	kfree(ring);
860 
861 free_ring_funcs:
862 	kfree(ring_funcs);
863 
864 	return r;
865 }
866