Linux Audio

Check our new training course

Loading...
v5.9
  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_shared.h"
 25
 26#include "amdgpu.h"
 27#include "amdgpu_gfx.h"
 28#include "amdgpu_dma_buf.h"
 29#include <linux/module.h>
 30#include <linux/dma-buf.h>
 31#include "amdgpu_xgmi.h"
 32#include <uapi/linux/kfd_ioctl.h>
 
 33
 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 */
 37uint64_t amdgpu_amdkfd_total_mem_size;
 38
 39int amdgpu_amdkfd_init(void)
 40{
 41	struct sysinfo si;
 42	int ret;
 43
 44	si_meminfo(&si);
 45	amdgpu_amdkfd_total_mem_size = si.totalram - si.totalhigh;
 46	amdgpu_amdkfd_total_mem_size *= si.mem_unit;
 47
 48#ifdef CONFIG_HSA_AMD
 49	ret = kgd2kfd_init();
 50	amdgpu_amdkfd_gpuvm_init_mem_limits();
 51#else
 52	ret = -ENOENT;
 53#endif
 54
 55	return ret;
 56}
 57
 58void amdgpu_amdkfd_fini(void)
 59{
 60	kgd2kfd_exit();
 61}
 62
 63void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
 64{
 65	bool vf = amdgpu_sriov_vf(adev);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 66
 67	adev->kfd.dev = kgd2kfd_probe((struct kgd_dev *)adev,
 68				      adev->pdev, adev->asic_type, vf);
 69
 70	if (adev->kfd.dev)
 71		amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
 72}
 73
 74/**
 75 * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
 76 *                                setup amdkfd
 77 *
 78 * @adev: amdgpu_device pointer
 79 * @aperture_base: output returning doorbell aperture base physical address
 80 * @aperture_size: output returning doorbell aperture size in bytes
 81 * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
 82 *
 83 * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
 84 * takes doorbells required for its own rings and reports the setup to amdkfd.
 85 * amdgpu reserved doorbells are at the start of the doorbell aperture.
 86 */
 87static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
 88					 phys_addr_t *aperture_base,
 89					 size_t *aperture_size,
 90					 size_t *start_offset)
 91{
 92	/*
 93	 * The first num_doorbells are used by amdgpu.
 94	 * amdkfd takes whatever's left in the aperture.
 95	 */
 96	if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) {
 97		*aperture_base = adev->doorbell.base;
 98		*aperture_size = adev->doorbell.size;
 99		*start_offset = adev->doorbell.num_doorbells * sizeof(u32);
100	} else {
101		*aperture_base = 0;
102		*aperture_size = 0;
103		*start_offset = 0;
104	}
105}
106
107void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
108{
109	int i;
110	int last_valid_bit;
111
112	if (adev->kfd.dev) {
113		struct kgd2kfd_shared_resources gpu_resources = {
114			.compute_vmid_bitmap =
115				((1 << AMDGPU_NUM_VMID) - 1) -
116				((1 << adev->vm_manager.first_kfd_vmid) - 1),
117			.num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
118			.num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe,
119			.gpuvm_size = min(adev->vm_manager.max_pfn
120					  << AMDGPU_GPU_PAGE_SHIFT,
121					  AMDGPU_GMC_HOLE_START),
122			.drm_render_minor = adev->ddev->render->index,
123			.sdma_doorbell_idx = adev->doorbell_index.sdma_engine,
124
125		};
126
127		/* this is going to have a few of the MSBs set that we need to
128		 * clear
129		 */
130		bitmap_complement(gpu_resources.cp_queue_bitmap,
131				  adev->gfx.mec.queue_bitmap,
132				  KGD_MAX_QUEUES);
133
 
 
 
 
 
 
 
 
134		/* According to linux/bitmap.h we shouldn't use bitmap_clear if
135		 * nbits is not compile time constant
136		 */
137		last_valid_bit = 1 /* only first MEC can have compute queues */
138				* adev->gfx.mec.num_pipe_per_mec
139				* adev->gfx.mec.num_queue_per_pipe;
140		for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i)
141			clear_bit(i, gpu_resources.cp_queue_bitmap);
142
143		amdgpu_doorbell_get_kfd_info(adev,
144				&gpu_resources.doorbell_physical_address,
145				&gpu_resources.doorbell_aperture_size,
146				&gpu_resources.doorbell_start_offset);
147
148		/* Since SOC15, BIF starts to statically use the
149		 * lower 12 bits of doorbell addresses for routing
150		 * based on settings in registers like
151		 * SDMA0_DOORBELL_RANGE etc..
152		 * In order to route a doorbell to CP engine, the lower
153		 * 12 bits of its address has to be outside the range
154		 * set for SDMA, VCN, and IH blocks.
155		 */
156		if (adev->asic_type >= CHIP_VEGA10) {
157			gpu_resources.non_cp_doorbells_start =
158					adev->doorbell_index.first_non_cp;
159			gpu_resources.non_cp_doorbells_end =
160					adev->doorbell_index.last_non_cp;
161		}
162
163		kgd2kfd_device_init(adev->kfd.dev, adev->ddev, &gpu_resources);
164	}
165}
166
167void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev)
168{
169	if (adev->kfd.dev) {
170		kgd2kfd_device_exit(adev->kfd.dev);
171		adev->kfd.dev = NULL;
172	}
173}
174
175void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
176		const void *ih_ring_entry)
177{
178	if (adev->kfd.dev)
179		kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
180}
181
182void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool run_pm)
183{
184	if (adev->kfd.dev)
185		kgd2kfd_suspend(adev->kfd.dev, run_pm);
186}
187
188int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool run_pm)
189{
190	int r = 0;
191
192	if (adev->kfd.dev)
193		r = kgd2kfd_resume(adev->kfd.dev, run_pm);
194
195	return r;
196}
197
198int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
199{
200	int r = 0;
201
202	if (adev->kfd.dev)
203		r = kgd2kfd_pre_reset(adev->kfd.dev);
204
205	return r;
206}
207
208int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
209{
210	int r = 0;
211
212	if (adev->kfd.dev)
213		r = kgd2kfd_post_reset(adev->kfd.dev);
214
215	return r;
216}
217
218void amdgpu_amdkfd_gpu_reset(struct kgd_dev *kgd)
219{
220	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
221
222	if (amdgpu_device_should_recover_gpu(adev))
223		amdgpu_device_gpu_recover(adev, NULL);
224}
225
226int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
227				void **mem_obj, uint64_t *gpu_addr,
228				void **cpu_ptr, bool cp_mqd_gfx9)
229{
230	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
231	struct amdgpu_bo *bo = NULL;
232	struct amdgpu_bo_param bp;
233	int r;
234	void *cpu_ptr_tmp = NULL;
235
236	memset(&bp, 0, sizeof(bp));
237	bp.size = size;
238	bp.byte_align = PAGE_SIZE;
239	bp.domain = AMDGPU_GEM_DOMAIN_GTT;
240	bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
241	bp.type = ttm_bo_type_kernel;
242	bp.resv = NULL;
243
244	if (cp_mqd_gfx9)
245		bp.flags |= AMDGPU_GEM_CREATE_CP_MQD_GFX9;
246
247	r = amdgpu_bo_create(adev, &bp, &bo);
248	if (r) {
249		dev_err(adev->dev,
250			"failed to allocate BO for amdkfd (%d)\n", r);
251		return r;
252	}
253
254	/* map the buffer */
255	r = amdgpu_bo_reserve(bo, true);
256	if (r) {
257		dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
258		goto allocate_mem_reserve_bo_failed;
259	}
260
261	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
262	if (r) {
263		dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r);
264		goto allocate_mem_pin_bo_failed;
265	}
266
267	r = amdgpu_ttm_alloc_gart(&bo->tbo);
268	if (r) {
269		dev_err(adev->dev, "%p bind failed\n", bo);
270		goto allocate_mem_kmap_bo_failed;
271	}
272
273	r = amdgpu_bo_kmap(bo, &cpu_ptr_tmp);
274	if (r) {
275		dev_err(adev->dev,
276			"(%d) failed to map bo to kernel for amdkfd\n", r);
277		goto allocate_mem_kmap_bo_failed;
278	}
279
280	*mem_obj = bo;
281	*gpu_addr = amdgpu_bo_gpu_offset(bo);
282	*cpu_ptr = cpu_ptr_tmp;
283
284	amdgpu_bo_unreserve(bo);
285
286	return 0;
287
288allocate_mem_kmap_bo_failed:
289	amdgpu_bo_unpin(bo);
290allocate_mem_pin_bo_failed:
291	amdgpu_bo_unreserve(bo);
292allocate_mem_reserve_bo_failed:
293	amdgpu_bo_unref(&bo);
294
295	return r;
296}
297
298void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
299{
300	struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
301
302	amdgpu_bo_reserve(bo, true);
303	amdgpu_bo_kunmap(bo);
304	amdgpu_bo_unpin(bo);
305	amdgpu_bo_unreserve(bo);
306	amdgpu_bo_unref(&(bo));
307}
308
309int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size,
310				void **mem_obj)
311{
312	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
313	struct amdgpu_bo *bo = NULL;
314	struct amdgpu_bo_param bp;
315	int r;
316
317	memset(&bp, 0, sizeof(bp));
318	bp.size = size;
319	bp.byte_align = 1;
320	bp.domain = AMDGPU_GEM_DOMAIN_GWS;
321	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
322	bp.type = ttm_bo_type_device;
323	bp.resv = NULL;
324
325	r = amdgpu_bo_create(adev, &bp, &bo);
326	if (r) {
327		dev_err(adev->dev,
328			"failed to allocate gws BO for amdkfd (%d)\n", r);
329		return r;
330	}
331
332	*mem_obj = bo;
333	return 0;
334}
335
336void amdgpu_amdkfd_free_gws(struct kgd_dev *kgd, void *mem_obj)
337{
338	struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj;
339
340	amdgpu_bo_unref(&bo);
341}
342
343uint32_t amdgpu_amdkfd_get_fw_version(struct kgd_dev *kgd,
344				      enum kgd_engine_type type)
345{
346	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
347
348	switch (type) {
349	case KGD_ENGINE_PFP:
350		return adev->gfx.pfp_fw_version;
351
352	case KGD_ENGINE_ME:
353		return adev->gfx.me_fw_version;
354
355	case KGD_ENGINE_CE:
356		return adev->gfx.ce_fw_version;
357
358	case KGD_ENGINE_MEC1:
359		return adev->gfx.mec_fw_version;
360
361	case KGD_ENGINE_MEC2:
362		return adev->gfx.mec2_fw_version;
363
364	case KGD_ENGINE_RLC:
365		return adev->gfx.rlc_fw_version;
366
367	case KGD_ENGINE_SDMA1:
368		return adev->sdma.instance[0].fw_version;
369
370	case KGD_ENGINE_SDMA2:
371		return adev->sdma.instance[1].fw_version;
372
373	default:
374		return 0;
375	}
376
377	return 0;
378}
379
380void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd,
381				      struct kfd_local_mem_info *mem_info)
382{
383	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
384	uint64_t address_mask = adev->dev->dma_mask ? ~*adev->dev->dma_mask :
385					     ~((1ULL << 32) - 1);
386	resource_size_t aper_limit = adev->gmc.aper_base + adev->gmc.aper_size;
387
388	memset(mem_info, 0, sizeof(*mem_info));
389	if (!(adev->gmc.aper_base & address_mask || aper_limit & address_mask)) {
390		mem_info->local_mem_size_public = adev->gmc.visible_vram_size;
391		mem_info->local_mem_size_private = adev->gmc.real_vram_size -
392				adev->gmc.visible_vram_size;
393	} else {
394		mem_info->local_mem_size_public = 0;
395		mem_info->local_mem_size_private = adev->gmc.real_vram_size;
396	}
397	mem_info->vram_width = adev->gmc.vram_width;
398
399	pr_debug("Address base: %pap limit %pap public 0x%llx private 0x%llx\n",
400			&adev->gmc.aper_base, &aper_limit,
401			mem_info->local_mem_size_public,
402			mem_info->local_mem_size_private);
403
404	if (amdgpu_sriov_vf(adev))
405		mem_info->mem_clk_max = adev->clock.default_mclk / 100;
406	else if (adev->pm.dpm_enabled) {
407		if (amdgpu_emu_mode == 1)
408			mem_info->mem_clk_max = 0;
409		else
410			mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
411	} else
412		mem_info->mem_clk_max = 100;
413}
414
415uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd)
416{
417	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
418
419	if (adev->gfx.funcs->get_gpu_clock_counter)
420		return adev->gfx.funcs->get_gpu_clock_counter(adev);
421	return 0;
422}
423
424uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
425{
426	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
427
428	/* the sclk is in quantas of 10kHz */
429	if (amdgpu_sriov_vf(adev))
430		return adev->clock.default_sclk / 100;
431	else if (adev->pm.dpm_enabled)
432		return amdgpu_dpm_get_sclk(adev, false) / 100;
433	else
434		return 100;
435}
436
437void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
438{
439	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
440	struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
441
442	memset(cu_info, 0, sizeof(*cu_info));
443	if (sizeof(cu_info->cu_bitmap) != sizeof(acu_info.bitmap))
444		return;
445
446	cu_info->cu_active_number = acu_info.number;
447	cu_info->cu_ao_mask = acu_info.ao_cu_mask;
448	memcpy(&cu_info->cu_bitmap[0], &acu_info.bitmap[0],
449	       sizeof(acu_info.bitmap));
450	cu_info->num_shader_engines = adev->gfx.config.max_shader_engines;
451	cu_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se;
452	cu_info->num_cu_per_sh = adev->gfx.config.max_cu_per_sh;
453	cu_info->simd_per_cu = acu_info.simd_per_cu;
454	cu_info->max_waves_per_simd = acu_info.max_waves_per_simd;
455	cu_info->wave_front_size = acu_info.wave_front_size;
456	cu_info->max_scratch_slots_per_cu = acu_info.max_scratch_slots_per_cu;
457	cu_info->lds_size = acu_info.lds_size;
458}
459
460int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,
461				  struct kgd_dev **dma_buf_kgd,
462				  uint64_t *bo_size, void *metadata_buffer,
463				  size_t buffer_size, uint32_t *metadata_size,
464				  uint32_t *flags)
465{
466	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
467	struct dma_buf *dma_buf;
468	struct drm_gem_object *obj;
469	struct amdgpu_bo *bo;
470	uint64_t metadata_flags;
471	int r = -EINVAL;
472
473	dma_buf = dma_buf_get(dma_buf_fd);
474	if (IS_ERR(dma_buf))
475		return PTR_ERR(dma_buf);
476
477	if (dma_buf->ops != &amdgpu_dmabuf_ops)
478		/* Can't handle non-graphics buffers */
479		goto out_put;
480
481	obj = dma_buf->priv;
482	if (obj->dev->driver != adev->ddev->driver)
483		/* Can't handle buffers from different drivers */
484		goto out_put;
485
486	adev = obj->dev->dev_private;
487	bo = gem_to_amdgpu_bo(obj);
488	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
489				    AMDGPU_GEM_DOMAIN_GTT)))
490		/* Only VRAM and GTT BOs are supported */
491		goto out_put;
492
493	r = 0;
494	if (dma_buf_kgd)
495		*dma_buf_kgd = (struct kgd_dev *)adev;
496	if (bo_size)
497		*bo_size = amdgpu_bo_size(bo);
498	if (metadata_size)
499		*metadata_size = bo->metadata_size;
500	if (metadata_buffer)
501		r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
502					   metadata_size, &metadata_flags);
503	if (flags) {
504		*flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
505				KFD_IOC_ALLOC_MEM_FLAGS_VRAM
506				: KFD_IOC_ALLOC_MEM_FLAGS_GTT;
507
508		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
509			*flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC;
510	}
511
512out_put:
513	dma_buf_put(dma_buf);
514	return r;
515}
516
517uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd)
518{
519	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
520
521	return amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
522}
523
524uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd)
525{
526	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
527
528	return adev->gmc.xgmi.hive_id;
529}
530
531uint64_t amdgpu_amdkfd_get_unique_id(struct kgd_dev *kgd)
532{
533	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
534
535	return adev->unique_id;
536}
537
538uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *src)
539{
540	struct amdgpu_device *peer_adev = (struct amdgpu_device *)src;
541	struct amdgpu_device *adev = (struct amdgpu_device *)dst;
542	int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev);
543
544	if (ret < 0) {
545		DRM_ERROR("amdgpu: failed to get  xgmi hops count between node %d and %d. ret = %d\n",
546			adev->gmc.xgmi.physical_node_id,
547			peer_adev->gmc.xgmi.physical_node_id, ret);
548		ret = 0;
549	}
550	return  (uint8_t)ret;
551}
552
553uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd)
554{
555	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
556
557	return adev->rmmio_remap.bus_addr;
558}
559
560uint32_t amdgpu_amdkfd_get_num_gws(struct kgd_dev *kgd)
561{
562	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
563
564	return adev->gds.gws_size;
565}
566
567uint32_t amdgpu_amdkfd_get_asic_rev_id(struct kgd_dev *kgd)
568{
569	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
570
571	return adev->rev_id;
572}
573
574int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
575				uint32_t vmid, uint64_t gpu_addr,
576				uint32_t *ib_cmd, uint32_t ib_len)
577{
578	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
579	struct amdgpu_job *job;
580	struct amdgpu_ib *ib;
581	struct amdgpu_ring *ring;
582	struct dma_fence *f = NULL;
583	int ret;
584
585	switch (engine) {
586	case KGD_ENGINE_MEC1:
587		ring = &adev->gfx.compute_ring[0];
588		break;
589	case KGD_ENGINE_SDMA1:
590		ring = &adev->sdma.instance[0].ring;
591		break;
592	case KGD_ENGINE_SDMA2:
593		ring = &adev->sdma.instance[1].ring;
594		break;
595	default:
596		pr_err("Invalid engine in IB submission: %d\n", engine);
597		ret = -EINVAL;
598		goto err;
599	}
600
601	ret = amdgpu_job_alloc(adev, 1, &job, NULL);
602	if (ret)
603		goto err;
604
605	ib = &job->ibs[0];
606	memset(ib, 0, sizeof(struct amdgpu_ib));
607
608	ib->gpu_addr = gpu_addr;
609	ib->ptr = ib_cmd;
610	ib->length_dw = ib_len;
611	/* This works for NO_HWS. TODO: need to handle without knowing VMID */
612	job->vmid = vmid;
613
614	ret = amdgpu_ib_schedule(ring, 1, ib, job, &f);
615	if (ret) {
616		DRM_ERROR("amdgpu: failed to schedule IB.\n");
617		goto err_ib_sched;
618	}
619
620	ret = dma_fence_wait(f, false);
621
622err_ib_sched:
623	dma_fence_put(f);
624	amdgpu_job_free(job);
625err:
626	return ret;
627}
628
629void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle)
630{
631	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
632
633	amdgpu_dpm_switch_power_profile(adev,
634					PP_SMC_POWER_PROFILE_COMPUTE,
635					!idle);
 
 
 
 
 
 
636}
637
638bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
639{
640	if (adev->kfd.dev)
641		return vmid >= adev->vm_manager.first_kfd_vmid;
642
643	return false;
644}
645
646int amdgpu_amdkfd_flush_gpu_tlb_vmid(struct kgd_dev *kgd, uint16_t vmid)
647{
648	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
649
650	if (adev->family == AMDGPU_FAMILY_AI) {
651		int i;
652
653		for (i = 0; i < adev->num_vmhubs; i++)
654			amdgpu_gmc_flush_gpu_tlb(adev, vmid, i, 0);
655	} else {
656		amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0);
657	}
658
659	return 0;
660}
661
662int amdgpu_amdkfd_flush_gpu_tlb_pasid(struct kgd_dev *kgd, uint16_t pasid)
663{
664	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
665	const uint32_t flush_type = 0;
666	bool all_hub = false;
667
668	if (adev->family == AMDGPU_FAMILY_AI)
669		all_hub = true;
670
671	return amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, flush_type, all_hub);
672}
673
674bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd)
675{
676	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
677
678	return adev->have_atomics_support;
679}
680
681#ifndef CONFIG_HSA_AMD
682bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
683{
684	return false;
685}
686
687void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
688{
689}
690
691int amdgpu_amdkfd_remove_fence_on_pt_pd_bos(struct amdgpu_bo *bo)
692{
693	return 0;
694}
695
696void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
697					struct amdgpu_vm *vm)
698{
699}
700
701struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
702{
703	return NULL;
704}
705
706int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm)
707{
708	return 0;
709}
710
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
711struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev,
712			      unsigned int asic_type, bool vf)
713{
714	return NULL;
715}
716
717bool kgd2kfd_device_init(struct kfd_dev *kfd,
718			 struct drm_device *ddev,
719			 const struct kgd2kfd_shared_resources *gpu_resources)
720{
721	return false;
722}
723
724void kgd2kfd_device_exit(struct kfd_dev *kfd)
725{
726}
727
728void kgd2kfd_exit(void)
729{
730}
731
732void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
733{
734}
735
736int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
737{
738	return 0;
739}
740
741int kgd2kfd_pre_reset(struct kfd_dev *kfd)
742{
743	return 0;
744}
745
746int kgd2kfd_post_reset(struct kfd_dev *kfd)
747{
748	return 0;
749}
750
751void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
752{
753}
754
755void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
756{
757}
758#endif
v5.4
  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_shared.h"
 25
 26#include "amdgpu.h"
 27#include "amdgpu_gfx.h"
 28#include "amdgpu_dma_buf.h"
 29#include <linux/module.h>
 30#include <linux/dma-buf.h>
 31#include "amdgpu_xgmi.h"
 32
 33static const unsigned int compute_vmid_bitmap = 0xFF00;
 34
 35/* Total memory size in system memory and all GPU VRAM. Used to
 36 * estimate worst case amount of memory to reserve for page tables
 37 */
 38uint64_t amdgpu_amdkfd_total_mem_size;
 39
 40int amdgpu_amdkfd_init(void)
 41{
 42	struct sysinfo si;
 43	int ret;
 44
 45	si_meminfo(&si);
 46	amdgpu_amdkfd_total_mem_size = si.totalram - si.totalhigh;
 47	amdgpu_amdkfd_total_mem_size *= si.mem_unit;
 48
 49#ifdef CONFIG_HSA_AMD
 50	ret = kgd2kfd_init();
 51	amdgpu_amdkfd_gpuvm_init_mem_limits();
 52#else
 53	ret = -ENOENT;
 54#endif
 55
 56	return ret;
 57}
 58
 59void amdgpu_amdkfd_fini(void)
 60{
 61	kgd2kfd_exit();
 62}
 63
 64void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
 65{
 66	const struct kfd2kgd_calls *kfd2kgd;
 67
 68	switch (adev->asic_type) {
 69#ifdef CONFIG_DRM_AMDGPU_CIK
 70	case CHIP_KAVERI:
 71	case CHIP_HAWAII:
 72		kfd2kgd = amdgpu_amdkfd_gfx_7_get_functions();
 73		break;
 74#endif
 75	case CHIP_CARRIZO:
 76	case CHIP_TONGA:
 77	case CHIP_FIJI:
 78	case CHIP_POLARIS10:
 79	case CHIP_POLARIS11:
 80	case CHIP_POLARIS12:
 81	case CHIP_VEGAM:
 82		kfd2kgd = amdgpu_amdkfd_gfx_8_0_get_functions();
 83		break;
 84	case CHIP_VEGA10:
 85	case CHIP_VEGA12:
 86	case CHIP_VEGA20:
 87	case CHIP_RAVEN:
 88		kfd2kgd = amdgpu_amdkfd_gfx_9_0_get_functions();
 89		break;
 90	case CHIP_ARCTURUS:
 91		kfd2kgd = amdgpu_amdkfd_arcturus_get_functions();
 92		break;
 93	case CHIP_NAVI10:
 94	case CHIP_NAVI14:
 95	case CHIP_NAVI12:
 96		kfd2kgd = amdgpu_amdkfd_gfx_10_0_get_functions();
 97		break;
 98	default:
 99		dev_info(adev->dev, "kfd not supported on this ASIC\n");
100		return;
101	}
102
103	adev->kfd.dev = kgd2kfd_probe((struct kgd_dev *)adev,
104				      adev->pdev, kfd2kgd);
105
106	if (adev->kfd.dev)
107		amdgpu_amdkfd_total_mem_size += adev->gmc.real_vram_size;
108}
109
110/**
111 * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
112 *                                setup amdkfd
113 *
114 * @adev: amdgpu_device pointer
115 * @aperture_base: output returning doorbell aperture base physical address
116 * @aperture_size: output returning doorbell aperture size in bytes
117 * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
118 *
119 * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
120 * takes doorbells required for its own rings and reports the setup to amdkfd.
121 * amdgpu reserved doorbells are at the start of the doorbell aperture.
122 */
123static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
124					 phys_addr_t *aperture_base,
125					 size_t *aperture_size,
126					 size_t *start_offset)
127{
128	/*
129	 * The first num_doorbells are used by amdgpu.
130	 * amdkfd takes whatever's left in the aperture.
131	 */
132	if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) {
133		*aperture_base = adev->doorbell.base;
134		*aperture_size = adev->doorbell.size;
135		*start_offset = adev->doorbell.num_doorbells * sizeof(u32);
136	} else {
137		*aperture_base = 0;
138		*aperture_size = 0;
139		*start_offset = 0;
140	}
141}
142
143void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
144{
145	int i;
146	int last_valid_bit;
147
148	if (adev->kfd.dev) {
149		struct kgd2kfd_shared_resources gpu_resources = {
150			.compute_vmid_bitmap = compute_vmid_bitmap,
 
 
151			.num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
152			.num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe,
153			.gpuvm_size = min(adev->vm_manager.max_pfn
154					  << AMDGPU_GPU_PAGE_SHIFT,
155					  AMDGPU_GMC_HOLE_START),
156			.drm_render_minor = adev->ddev->render->index,
157			.sdma_doorbell_idx = adev->doorbell_index.sdma_engine,
158
159		};
160
161		/* this is going to have a few of the MSBs set that we need to
162		 * clear
163		 */
164		bitmap_complement(gpu_resources.queue_bitmap,
165				  adev->gfx.mec.queue_bitmap,
166				  KGD_MAX_QUEUES);
167
168		/* remove the KIQ bit as well */
169		if (adev->gfx.kiq.ring.sched.ready)
170			clear_bit(amdgpu_gfx_mec_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
176		/* According to linux/bitmap.h we shouldn't use bitmap_clear if
177		 * nbits is not compile time constant
178		 */
179		last_valid_bit = 1 /* only first MEC can have compute queues */
180				* adev->gfx.mec.num_pipe_per_mec
181				* adev->gfx.mec.num_queue_per_pipe;
182		for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i)
183			clear_bit(i, gpu_resources.queue_bitmap);
184
185		amdgpu_doorbell_get_kfd_info(adev,
186				&gpu_resources.doorbell_physical_address,
187				&gpu_resources.doorbell_aperture_size,
188				&gpu_resources.doorbell_start_offset);
189
190		/* Since SOC15, BIF starts to statically use the
191		 * lower 12 bits of doorbell addresses for routing
192		 * based on settings in registers like
193		 * SDMA0_DOORBELL_RANGE etc..
194		 * In order to route a doorbell to CP engine, the lower
195		 * 12 bits of its address has to be outside the range
196		 * set for SDMA, VCN, and IH blocks.
197		 */
198		if (adev->asic_type >= CHIP_VEGA10) {
199			gpu_resources.non_cp_doorbells_start =
200					adev->doorbell_index.first_non_cp;
201			gpu_resources.non_cp_doorbells_end =
202					adev->doorbell_index.last_non_cp;
203		}
204
205		kgd2kfd_device_init(adev->kfd.dev, &gpu_resources);
206	}
207}
208
209void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev)
210{
211	if (adev->kfd.dev) {
212		kgd2kfd_device_exit(adev->kfd.dev);
213		adev->kfd.dev = NULL;
214	}
215}
216
217void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
218		const void *ih_ring_entry)
219{
220	if (adev->kfd.dev)
221		kgd2kfd_interrupt(adev->kfd.dev, ih_ring_entry);
222}
223
224void amdgpu_amdkfd_suspend(struct amdgpu_device *adev)
225{
226	if (adev->kfd.dev)
227		kgd2kfd_suspend(adev->kfd.dev);
228}
229
230int amdgpu_amdkfd_resume(struct amdgpu_device *adev)
231{
232	int r = 0;
233
234	if (adev->kfd.dev)
235		r = kgd2kfd_resume(adev->kfd.dev);
236
237	return r;
238}
239
240int amdgpu_amdkfd_pre_reset(struct amdgpu_device *adev)
241{
242	int r = 0;
243
244	if (adev->kfd.dev)
245		r = kgd2kfd_pre_reset(adev->kfd.dev);
246
247	return r;
248}
249
250int amdgpu_amdkfd_post_reset(struct amdgpu_device *adev)
251{
252	int r = 0;
253
254	if (adev->kfd.dev)
255		r = kgd2kfd_post_reset(adev->kfd.dev);
256
257	return r;
258}
259
260void amdgpu_amdkfd_gpu_reset(struct kgd_dev *kgd)
261{
262	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
263
264	if (amdgpu_device_should_recover_gpu(adev))
265		amdgpu_device_gpu_recover(adev, NULL);
266}
267
268int amdgpu_amdkfd_alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
269				void **mem_obj, uint64_t *gpu_addr,
270				void **cpu_ptr, bool mqd_gfx9)
271{
272	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
273	struct amdgpu_bo *bo = NULL;
274	struct amdgpu_bo_param bp;
275	int r;
276	void *cpu_ptr_tmp = NULL;
277
278	memset(&bp, 0, sizeof(bp));
279	bp.size = size;
280	bp.byte_align = PAGE_SIZE;
281	bp.domain = AMDGPU_GEM_DOMAIN_GTT;
282	bp.flags = AMDGPU_GEM_CREATE_CPU_GTT_USWC;
283	bp.type = ttm_bo_type_kernel;
284	bp.resv = NULL;
285
286	if (mqd_gfx9)
287		bp.flags |= AMDGPU_GEM_CREATE_MQD_GFX9;
288
289	r = amdgpu_bo_create(adev, &bp, &bo);
290	if (r) {
291		dev_err(adev->dev,
292			"failed to allocate BO for amdkfd (%d)\n", r);
293		return r;
294	}
295
296	/* map the buffer */
297	r = amdgpu_bo_reserve(bo, true);
298	if (r) {
299		dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
300		goto allocate_mem_reserve_bo_failed;
301	}
302
303	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
304	if (r) {
305		dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r);
306		goto allocate_mem_pin_bo_failed;
307	}
308
309	r = amdgpu_ttm_alloc_gart(&bo->tbo);
310	if (r) {
311		dev_err(adev->dev, "%p bind failed\n", bo);
312		goto allocate_mem_kmap_bo_failed;
313	}
314
315	r = amdgpu_bo_kmap(bo, &cpu_ptr_tmp);
316	if (r) {
317		dev_err(adev->dev,
318			"(%d) failed to map bo to kernel for amdkfd\n", r);
319		goto allocate_mem_kmap_bo_failed;
320	}
321
322	*mem_obj = bo;
323	*gpu_addr = amdgpu_bo_gpu_offset(bo);
324	*cpu_ptr = cpu_ptr_tmp;
325
326	amdgpu_bo_unreserve(bo);
327
328	return 0;
329
330allocate_mem_kmap_bo_failed:
331	amdgpu_bo_unpin(bo);
332allocate_mem_pin_bo_failed:
333	amdgpu_bo_unreserve(bo);
334allocate_mem_reserve_bo_failed:
335	amdgpu_bo_unref(&bo);
336
337	return r;
338}
339
340void amdgpu_amdkfd_free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
341{
342	struct amdgpu_bo *bo = (struct amdgpu_bo *) mem_obj;
343
344	amdgpu_bo_reserve(bo, true);
345	amdgpu_bo_kunmap(bo);
346	amdgpu_bo_unpin(bo);
347	amdgpu_bo_unreserve(bo);
348	amdgpu_bo_unref(&(bo));
349}
350
351int amdgpu_amdkfd_alloc_gws(struct kgd_dev *kgd, size_t size,
352				void **mem_obj)
353{
354	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
355	struct amdgpu_bo *bo = NULL;
356	struct amdgpu_bo_param bp;
357	int r;
358
359	memset(&bp, 0, sizeof(bp));
360	bp.size = size;
361	bp.byte_align = 1;
362	bp.domain = AMDGPU_GEM_DOMAIN_GWS;
363	bp.flags = AMDGPU_GEM_CREATE_NO_CPU_ACCESS;
364	bp.type = ttm_bo_type_device;
365	bp.resv = NULL;
366
367	r = amdgpu_bo_create(adev, &bp, &bo);
368	if (r) {
369		dev_err(adev->dev,
370			"failed to allocate gws BO for amdkfd (%d)\n", r);
371		return r;
372	}
373
374	*mem_obj = bo;
375	return 0;
376}
377
378void amdgpu_amdkfd_free_gws(struct kgd_dev *kgd, void *mem_obj)
379{
380	struct amdgpu_bo *bo = (struct amdgpu_bo *)mem_obj;
381
382	amdgpu_bo_unref(&bo);
383}
384
385uint32_t amdgpu_amdkfd_get_fw_version(struct kgd_dev *kgd,
386				      enum kgd_engine_type type)
387{
388	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
389
390	switch (type) {
391	case KGD_ENGINE_PFP:
392		return adev->gfx.pfp_fw_version;
393
394	case KGD_ENGINE_ME:
395		return adev->gfx.me_fw_version;
396
397	case KGD_ENGINE_CE:
398		return adev->gfx.ce_fw_version;
399
400	case KGD_ENGINE_MEC1:
401		return adev->gfx.mec_fw_version;
402
403	case KGD_ENGINE_MEC2:
404		return adev->gfx.mec2_fw_version;
405
406	case KGD_ENGINE_RLC:
407		return adev->gfx.rlc_fw_version;
408
409	case KGD_ENGINE_SDMA1:
410		return adev->sdma.instance[0].fw_version;
411
412	case KGD_ENGINE_SDMA2:
413		return adev->sdma.instance[1].fw_version;
414
415	default:
416		return 0;
417	}
418
419	return 0;
420}
421
422void amdgpu_amdkfd_get_local_mem_info(struct kgd_dev *kgd,
423				      struct kfd_local_mem_info *mem_info)
424{
425	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
426	uint64_t address_mask = adev->dev->dma_mask ? ~*adev->dev->dma_mask :
427					     ~((1ULL << 32) - 1);
428	resource_size_t aper_limit = adev->gmc.aper_base + adev->gmc.aper_size;
429
430	memset(mem_info, 0, sizeof(*mem_info));
431	if (!(adev->gmc.aper_base & address_mask || aper_limit & address_mask)) {
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	} else {
436		mem_info->local_mem_size_public = 0;
437		mem_info->local_mem_size_private = adev->gmc.real_vram_size;
438	}
439	mem_info->vram_width = adev->gmc.vram_width;
440
441	pr_debug("Address base: %pap limit %pap public 0x%llx private 0x%llx\n",
442			&adev->gmc.aper_base, &aper_limit,
443			mem_info->local_mem_size_public,
444			mem_info->local_mem_size_private);
445
446	if (amdgpu_sriov_vf(adev))
447		mem_info->mem_clk_max = adev->clock.default_mclk / 100;
448	else if (adev->powerplay.pp_funcs) {
449		if (amdgpu_emu_mode == 1)
450			mem_info->mem_clk_max = 0;
451		else
452			mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
453	} else
454		mem_info->mem_clk_max = 100;
455}
456
457uint64_t amdgpu_amdkfd_get_gpu_clock_counter(struct kgd_dev *kgd)
458{
459	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
460
461	if (adev->gfx.funcs->get_gpu_clock_counter)
462		return adev->gfx.funcs->get_gpu_clock_counter(adev);
463	return 0;
464}
465
466uint32_t amdgpu_amdkfd_get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
467{
468	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
469
470	/* the sclk is in quantas of 10kHz */
471	if (amdgpu_sriov_vf(adev))
472		return adev->clock.default_sclk / 100;
473	else if (adev->powerplay.pp_funcs)
474		return amdgpu_dpm_get_sclk(adev, false) / 100;
475	else
476		return 100;
477}
478
479void amdgpu_amdkfd_get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
480{
481	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
482	struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
483
484	memset(cu_info, 0, sizeof(*cu_info));
485	if (sizeof(cu_info->cu_bitmap) != sizeof(acu_info.bitmap))
486		return;
487
488	cu_info->cu_active_number = acu_info.number;
489	cu_info->cu_ao_mask = acu_info.ao_cu_mask;
490	memcpy(&cu_info->cu_bitmap[0], &acu_info.bitmap[0],
491	       sizeof(acu_info.bitmap));
492	cu_info->num_shader_engines = adev->gfx.config.max_shader_engines;
493	cu_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se;
494	cu_info->num_cu_per_sh = adev->gfx.config.max_cu_per_sh;
495	cu_info->simd_per_cu = acu_info.simd_per_cu;
496	cu_info->max_waves_per_simd = acu_info.max_waves_per_simd;
497	cu_info->wave_front_size = acu_info.wave_front_size;
498	cu_info->max_scratch_slots_per_cu = acu_info.max_scratch_slots_per_cu;
499	cu_info->lds_size = acu_info.lds_size;
500}
501
502int amdgpu_amdkfd_get_dmabuf_info(struct kgd_dev *kgd, int dma_buf_fd,
503				  struct kgd_dev **dma_buf_kgd,
504				  uint64_t *bo_size, void *metadata_buffer,
505				  size_t buffer_size, uint32_t *metadata_size,
506				  uint32_t *flags)
507{
508	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
509	struct dma_buf *dma_buf;
510	struct drm_gem_object *obj;
511	struct amdgpu_bo *bo;
512	uint64_t metadata_flags;
513	int r = -EINVAL;
514
515	dma_buf = dma_buf_get(dma_buf_fd);
516	if (IS_ERR(dma_buf))
517		return PTR_ERR(dma_buf);
518
519	if (dma_buf->ops != &amdgpu_dmabuf_ops)
520		/* Can't handle non-graphics buffers */
521		goto out_put;
522
523	obj = dma_buf->priv;
524	if (obj->dev->driver != adev->ddev->driver)
525		/* Can't handle buffers from different drivers */
526		goto out_put;
527
528	adev = obj->dev->dev_private;
529	bo = gem_to_amdgpu_bo(obj);
530	if (!(bo->preferred_domains & (AMDGPU_GEM_DOMAIN_VRAM |
531				    AMDGPU_GEM_DOMAIN_GTT)))
532		/* Only VRAM and GTT BOs are supported */
533		goto out_put;
534
535	r = 0;
536	if (dma_buf_kgd)
537		*dma_buf_kgd = (struct kgd_dev *)adev;
538	if (bo_size)
539		*bo_size = amdgpu_bo_size(bo);
540	if (metadata_size)
541		*metadata_size = bo->metadata_size;
542	if (metadata_buffer)
543		r = amdgpu_bo_get_metadata(bo, metadata_buffer, buffer_size,
544					   metadata_size, &metadata_flags);
545	if (flags) {
546		*flags = (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) ?
547			ALLOC_MEM_FLAGS_VRAM : ALLOC_MEM_FLAGS_GTT;
 
548
549		if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
550			*flags |= ALLOC_MEM_FLAGS_PUBLIC;
551	}
552
553out_put:
554	dma_buf_put(dma_buf);
555	return r;
556}
557
558uint64_t amdgpu_amdkfd_get_vram_usage(struct kgd_dev *kgd)
559{
560	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
561
562	return amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
563}
564
565uint64_t amdgpu_amdkfd_get_hive_id(struct kgd_dev *kgd)
566{
567	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
568
569	return adev->gmc.xgmi.hive_id;
570}
 
 
 
 
 
 
 
 
571uint8_t amdgpu_amdkfd_get_xgmi_hops_count(struct kgd_dev *dst, struct kgd_dev *src)
572{
573	struct amdgpu_device *peer_adev = (struct amdgpu_device *)src;
574	struct amdgpu_device *adev = (struct amdgpu_device *)dst;
575	int ret = amdgpu_xgmi_get_hops_count(adev, peer_adev);
576
577	if (ret < 0) {
578		DRM_ERROR("amdgpu: failed to get  xgmi hops count between node %d and %d. ret = %d\n",
579			adev->gmc.xgmi.physical_node_id,
580			peer_adev->gmc.xgmi.physical_node_id, ret);
581		ret = 0;
582	}
583	return  (uint8_t)ret;
584}
585
586uint64_t amdgpu_amdkfd_get_mmio_remap_phys_addr(struct kgd_dev *kgd)
587{
588	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
589
590	return adev->rmmio_remap.bus_addr;
591}
592
593uint32_t amdgpu_amdkfd_get_num_gws(struct kgd_dev *kgd)
594{
595	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
596
597	return adev->gds.gws_size;
598}
599
 
 
 
 
 
 
 
600int amdgpu_amdkfd_submit_ib(struct kgd_dev *kgd, enum kgd_engine_type engine,
601				uint32_t vmid, uint64_t gpu_addr,
602				uint32_t *ib_cmd, uint32_t ib_len)
603{
604	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
605	struct amdgpu_job *job;
606	struct amdgpu_ib *ib;
607	struct amdgpu_ring *ring;
608	struct dma_fence *f = NULL;
609	int ret;
610
611	switch (engine) {
612	case KGD_ENGINE_MEC1:
613		ring = &adev->gfx.compute_ring[0];
614		break;
615	case KGD_ENGINE_SDMA1:
616		ring = &adev->sdma.instance[0].ring;
617		break;
618	case KGD_ENGINE_SDMA2:
619		ring = &adev->sdma.instance[1].ring;
620		break;
621	default:
622		pr_err("Invalid engine in IB submission: %d\n", engine);
623		ret = -EINVAL;
624		goto err;
625	}
626
627	ret = amdgpu_job_alloc(adev, 1, &job, NULL);
628	if (ret)
629		goto err;
630
631	ib = &job->ibs[0];
632	memset(ib, 0, sizeof(struct amdgpu_ib));
633
634	ib->gpu_addr = gpu_addr;
635	ib->ptr = ib_cmd;
636	ib->length_dw = ib_len;
637	/* This works for NO_HWS. TODO: need to handle without knowing VMID */
638	job->vmid = vmid;
639
640	ret = amdgpu_ib_schedule(ring, 1, ib, job, &f);
641	if (ret) {
642		DRM_ERROR("amdgpu: failed to schedule IB.\n");
643		goto err_ib_sched;
644	}
645
646	ret = dma_fence_wait(f, false);
647
648err_ib_sched:
649	dma_fence_put(f);
650	amdgpu_job_free(job);
651err:
652	return ret;
653}
654
655void amdgpu_amdkfd_set_compute_idle(struct kgd_dev *kgd, bool idle)
656{
657	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
658
659	if (is_support_sw_smu(adev))
660		smu_switch_power_profile(&adev->smu,
661					 PP_SMC_POWER_PROFILE_COMPUTE,
662					 !idle);
663	else if (adev->powerplay.pp_funcs &&
664		 adev->powerplay.pp_funcs->switch_power_profile)
665		amdgpu_dpm_switch_power_profile(adev,
666						PP_SMC_POWER_PROFILE_COMPUTE,
667						!idle);
668}
669
670bool amdgpu_amdkfd_is_kfd_vmid(struct amdgpu_device *adev, u32 vmid)
671{
672	if (adev->kfd.dev) {
673		if ((1 << vmid) & compute_vmid_bitmap)
674			return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
675	}
676
677	return false;
 
 
 
 
 
 
 
 
 
 
 
 
678}
679
680bool amdgpu_amdkfd_have_atomics_support(struct kgd_dev *kgd)
681{
682	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
683
684	return adev->have_atomics_support;
685}
686
687#ifndef CONFIG_HSA_AMD
688bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
689{
690	return false;
691}
692
693void amdgpu_amdkfd_unreserve_memory_limit(struct amdgpu_bo *bo)
694{
695}
696
 
 
 
 
 
697void amdgpu_amdkfd_gpuvm_destroy_cb(struct amdgpu_device *adev,
698					struct amdgpu_vm *vm)
699{
700}
701
702struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
703{
704	return NULL;
705}
706
707int amdgpu_amdkfd_evict_userptr(struct kgd_mem *mem, struct mm_struct *mm)
708{
709	return 0;
710}
711
712struct kfd2kgd_calls *amdgpu_amdkfd_gfx_7_get_functions(void)
713{
714	return NULL;
715}
716
717struct kfd2kgd_calls *amdgpu_amdkfd_gfx_8_0_get_functions(void)
718{
719	return NULL;
720}
721
722struct kfd2kgd_calls *amdgpu_amdkfd_gfx_9_0_get_functions(void)
723{
724	return NULL;
725}
726
727struct kfd2kgd_calls *amdgpu_amdkfd_arcturus_get_functions(void)
728{
729	return NULL;
730}
731
732struct kfd2kgd_calls *amdgpu_amdkfd_gfx_10_0_get_functions(void)
733{
734	return NULL;
735}
736
737struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev,
738			      const struct kfd2kgd_calls *f2g)
739{
740	return NULL;
741}
742
743bool kgd2kfd_device_init(struct kfd_dev *kfd,
 
744			 const struct kgd2kfd_shared_resources *gpu_resources)
745{
746	return false;
747}
748
749void kgd2kfd_device_exit(struct kfd_dev *kfd)
750{
751}
752
753void kgd2kfd_exit(void)
754{
755}
756
757void kgd2kfd_suspend(struct kfd_dev *kfd)
758{
759}
760
761int kgd2kfd_resume(struct kfd_dev *kfd)
762{
763	return 0;
764}
765
766int kgd2kfd_pre_reset(struct kfd_dev *kfd)
767{
768	return 0;
769}
770
771int kgd2kfd_post_reset(struct kfd_dev *kfd)
772{
773	return 0;
774}
775
776void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
777{
778}
779
780void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
781{
782}
783#endif