xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c (revision 1b392348de8ffc340545c62079645cb4695f5602)
1 /*
2  * Copyright 2022 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.h"
24 #include "amdgpu_xcp.h"
25 #include "amdgpu_drv.h"
26 
27 #include <drm/drm_drv.h>
28 #include "../amdxcp/amdgpu_xcp_drv.h"
29 
30 static void amdgpu_xcp_sysfs_entries_init(struct amdgpu_xcp_mgr *xcp_mgr);
31 static void amdgpu_xcp_sysfs_entries_update(struct amdgpu_xcp_mgr *xcp_mgr);
32 
33 static int __amdgpu_xcp_run(struct amdgpu_xcp_mgr *xcp_mgr,
34 			    struct amdgpu_xcp_ip *xcp_ip, int xcp_state)
35 {
36 	int (*run_func)(void *handle, uint32_t inst_mask);
37 	int ret = 0;
38 
39 	if (!xcp_ip || !xcp_ip->valid || !xcp_ip->ip_funcs)
40 		return 0;
41 
42 	run_func = NULL;
43 
44 	switch (xcp_state) {
45 	case AMDGPU_XCP_PREPARE_SUSPEND:
46 		run_func = xcp_ip->ip_funcs->prepare_suspend;
47 		break;
48 	case AMDGPU_XCP_SUSPEND:
49 		run_func = xcp_ip->ip_funcs->suspend;
50 		break;
51 	case AMDGPU_XCP_PREPARE_RESUME:
52 		run_func = xcp_ip->ip_funcs->prepare_resume;
53 		break;
54 	case AMDGPU_XCP_RESUME:
55 		run_func = xcp_ip->ip_funcs->resume;
56 		break;
57 	}
58 
59 	if (run_func)
60 		ret = run_func(xcp_mgr->adev, xcp_ip->inst_mask);
61 
62 	return ret;
63 }
64 
65 static int amdgpu_xcp_run_transition(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id,
66 				     int state)
67 {
68 	struct amdgpu_xcp_ip *xcp_ip;
69 	struct amdgpu_xcp *xcp;
70 	int i, ret;
71 
72 	if (xcp_id >= MAX_XCP || !xcp_mgr->xcp[xcp_id].valid)
73 		return -EINVAL;
74 
75 	xcp = &xcp_mgr->xcp[xcp_id];
76 	for (i = 0; i < AMDGPU_XCP_MAX_BLOCKS; ++i) {
77 		xcp_ip = &xcp->ip[i];
78 		ret = __amdgpu_xcp_run(xcp_mgr, xcp_ip, state);
79 		if (ret)
80 			break;
81 	}
82 
83 	return ret;
84 }
85 
86 int amdgpu_xcp_prepare_suspend(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id)
87 {
88 	return amdgpu_xcp_run_transition(xcp_mgr, xcp_id,
89 					 AMDGPU_XCP_PREPARE_SUSPEND);
90 }
91 
92 int amdgpu_xcp_suspend(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id)
93 {
94 	return amdgpu_xcp_run_transition(xcp_mgr, xcp_id, AMDGPU_XCP_SUSPEND);
95 }
96 
97 int amdgpu_xcp_prepare_resume(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id)
98 {
99 	return amdgpu_xcp_run_transition(xcp_mgr, xcp_id,
100 					 AMDGPU_XCP_PREPARE_RESUME);
101 }
102 
103 int amdgpu_xcp_resume(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id)
104 {
105 	return amdgpu_xcp_run_transition(xcp_mgr, xcp_id, AMDGPU_XCP_RESUME);
106 }
107 
108 static void __amdgpu_xcp_add_block(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id,
109 				   struct amdgpu_xcp_ip *ip)
110 {
111 	struct amdgpu_xcp *xcp;
112 
113 	if (!ip)
114 		return;
115 
116 	xcp = &xcp_mgr->xcp[xcp_id];
117 	xcp->ip[ip->ip_id] = *ip;
118 	xcp->ip[ip->ip_id].valid = true;
119 
120 	xcp->valid = true;
121 }
122 
123 int amdgpu_xcp_init(struct amdgpu_xcp_mgr *xcp_mgr, int num_xcps, int mode)
124 {
125 	struct amdgpu_device *adev = xcp_mgr->adev;
126 	struct amdgpu_xcp_ip ip;
127 	uint8_t mem_id;
128 	int i, j, ret;
129 
130 	if (!num_xcps || num_xcps > MAX_XCP)
131 		return -EINVAL;
132 
133 	xcp_mgr->mode = mode;
134 
135 	for (i = 0; i < MAX_XCP; ++i)
136 		xcp_mgr->xcp[i].valid = false;
137 
138 	/* This is needed for figuring out memory id of xcp */
139 	xcp_mgr->num_xcp_per_mem_partition = num_xcps / xcp_mgr->adev->gmc.num_mem_partitions;
140 
141 	for (i = 0; i < num_xcps; ++i) {
142 		for (j = AMDGPU_XCP_GFXHUB; j < AMDGPU_XCP_MAX_BLOCKS; ++j) {
143 			ret = xcp_mgr->funcs->get_ip_details(xcp_mgr, i, j,
144 							     &ip);
145 			if (ret)
146 				continue;
147 
148 			__amdgpu_xcp_add_block(xcp_mgr, i, &ip);
149 		}
150 
151 		xcp_mgr->xcp[i].id = i;
152 
153 		if (xcp_mgr->funcs->get_xcp_mem_id) {
154 			ret = xcp_mgr->funcs->get_xcp_mem_id(
155 				xcp_mgr, &xcp_mgr->xcp[i], &mem_id);
156 			if (ret)
157 				continue;
158 			else
159 				xcp_mgr->xcp[i].mem_id = mem_id;
160 		}
161 	}
162 
163 	xcp_mgr->num_xcps = num_xcps;
164 	amdgpu_xcp_update_partition_sched_list(adev);
165 
166 	return 0;
167 }
168 
169 static int __amdgpu_xcp_switch_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr,
170 					      int mode)
171 {
172 	int ret, curr_mode, num_xcps = 0;
173 
174 	if (!xcp_mgr->funcs || !xcp_mgr->funcs->switch_partition_mode)
175 		return 0;
176 
177 	mutex_lock(&xcp_mgr->xcp_lock);
178 
179 	curr_mode = xcp_mgr->mode;
180 	/* State set to transient mode */
181 	xcp_mgr->mode = AMDGPU_XCP_MODE_TRANS;
182 
183 	ret = xcp_mgr->funcs->switch_partition_mode(xcp_mgr, mode, &num_xcps);
184 
185 	if (ret) {
186 		/* Failed, get whatever mode it's at now */
187 		if (xcp_mgr->funcs->query_partition_mode)
188 			xcp_mgr->mode = amdgpu_xcp_query_partition_mode(
189 				xcp_mgr, AMDGPU_XCP_FL_LOCKED);
190 		else
191 			xcp_mgr->mode = curr_mode;
192 
193 		goto out;
194 	}
195 	amdgpu_xcp_sysfs_entries_update(xcp_mgr);
196 out:
197 	mutex_unlock(&xcp_mgr->xcp_lock);
198 
199 	return ret;
200 }
201 
202 int amdgpu_xcp_switch_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, int mode)
203 {
204 	if (!xcp_mgr || mode == AMDGPU_XCP_MODE_NONE)
205 		return -EINVAL;
206 
207 	if (xcp_mgr->mode == mode)
208 		return 0;
209 
210 	return __amdgpu_xcp_switch_partition_mode(xcp_mgr, mode);
211 }
212 
213 int amdgpu_xcp_restore_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr)
214 {
215 	if (!xcp_mgr || xcp_mgr->mode == AMDGPU_XCP_MODE_NONE)
216 		return 0;
217 
218 	return __amdgpu_xcp_switch_partition_mode(xcp_mgr, xcp_mgr->mode);
219 }
220 
221 static bool __amdgpu_xcp_is_cached_mode_valid(struct amdgpu_xcp_mgr *xcp_mgr)
222 {
223 	if (!xcp_mgr->funcs || !xcp_mgr->funcs->query_partition_mode)
224 		return true;
225 
226 	if (!amdgpu_sriov_vf(xcp_mgr->adev) &&
227 	    xcp_mgr->mode == AMDGPU_XCP_MODE_NONE)
228 		return true;
229 
230 	if (xcp_mgr->mode != AMDGPU_XCP_MODE_NONE &&
231 	    xcp_mgr->mode != AMDGPU_XCP_MODE_TRANS)
232 		return true;
233 
234 	return false;
235 }
236 
237 int amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, u32 flags)
238 {
239 	int mode;
240 
241 	if (__amdgpu_xcp_is_cached_mode_valid(xcp_mgr))
242 		return xcp_mgr->mode;
243 
244 	if (!(flags & AMDGPU_XCP_FL_LOCKED))
245 		mutex_lock(&xcp_mgr->xcp_lock);
246 	mode = xcp_mgr->funcs->query_partition_mode(xcp_mgr);
247 
248 	/* First time query for VF, set the mode here */
249 	if (amdgpu_sriov_vf(xcp_mgr->adev) &&
250 	    xcp_mgr->mode == AMDGPU_XCP_MODE_NONE)
251 		xcp_mgr->mode = mode;
252 
253 	if (xcp_mgr->mode != AMDGPU_XCP_MODE_TRANS && mode != xcp_mgr->mode)
254 		dev_WARN(
255 			xcp_mgr->adev->dev,
256 			"Cached partition mode %d not matching with device mode %d",
257 			xcp_mgr->mode, mode);
258 
259 	if (!(flags & AMDGPU_XCP_FL_LOCKED))
260 		mutex_unlock(&xcp_mgr->xcp_lock);
261 
262 	return mode;
263 }
264 
265 static int amdgpu_xcp_dev_alloc(struct amdgpu_device *adev)
266 {
267 	struct drm_device *p_ddev;
268 	struct drm_device *ddev;
269 	int i, ret;
270 
271 	ddev = adev_to_drm(adev);
272 
273 	/* xcp #0 shares drm device setting with adev */
274 	adev->xcp_mgr->xcp->ddev = ddev;
275 
276 	for (i = 1; i < MAX_XCP; i++) {
277 		ret = amdgpu_xcp_drm_dev_alloc(&p_ddev);
278 		if (ret == -ENOSPC) {
279 			dev_warn(adev->dev,
280 			"Skip xcp node #%d when out of drm node resource.", i);
281 			ret = 0;
282 			goto out;
283 		} else if (ret) {
284 			goto out;
285 		}
286 
287 		/* Redirect all IOCTLs to the primary device */
288 		adev->xcp_mgr->xcp[i].rdev = p_ddev->render->dev;
289 		adev->xcp_mgr->xcp[i].pdev = p_ddev->primary->dev;
290 		adev->xcp_mgr->xcp[i].driver = (struct drm_driver *)p_ddev->driver;
291 		adev->xcp_mgr->xcp[i].vma_offset_manager = p_ddev->vma_offset_manager;
292 		p_ddev->render->dev = ddev;
293 		p_ddev->primary->dev = ddev;
294 		p_ddev->vma_offset_manager = ddev->vma_offset_manager;
295 		p_ddev->driver = &amdgpu_partition_driver;
296 		adev->xcp_mgr->xcp[i].ddev = p_ddev;
297 
298 		dev_set_drvdata(p_ddev->dev, &adev->xcp_mgr->xcp[i]);
299 	}
300 	ret = 0;
301 out:
302 	amdgpu_xcp_sysfs_entries_init(adev->xcp_mgr);
303 
304 	return ret;
305 }
306 
307 int amdgpu_xcp_mgr_init(struct amdgpu_device *adev, int init_mode,
308 			int init_num_xcps,
309 			struct amdgpu_xcp_mgr_funcs *xcp_funcs)
310 {
311 	struct amdgpu_xcp_mgr *xcp_mgr;
312 	int i;
313 
314 	if (!xcp_funcs || !xcp_funcs->get_ip_details)
315 		return -EINVAL;
316 
317 	xcp_mgr = kzalloc(sizeof(*xcp_mgr), GFP_KERNEL);
318 
319 	if (!xcp_mgr)
320 		return -ENOMEM;
321 
322 	xcp_mgr->adev = adev;
323 	xcp_mgr->funcs = xcp_funcs;
324 	xcp_mgr->mode = init_mode;
325 	mutex_init(&xcp_mgr->xcp_lock);
326 
327 	if (init_mode != AMDGPU_XCP_MODE_NONE)
328 		amdgpu_xcp_init(xcp_mgr, init_num_xcps, init_mode);
329 
330 	adev->xcp_mgr = xcp_mgr;
331 	for (i = 0; i < MAX_XCP; ++i)
332 		xcp_mgr->xcp[i].xcp_mgr = xcp_mgr;
333 
334 	return amdgpu_xcp_dev_alloc(adev);
335 }
336 
337 int amdgpu_xcp_get_partition(struct amdgpu_xcp_mgr *xcp_mgr,
338 			     enum AMDGPU_XCP_IP_BLOCK ip, int instance)
339 {
340 	struct amdgpu_xcp *xcp;
341 	int i, id_mask = 0;
342 
343 	if (ip >= AMDGPU_XCP_MAX_BLOCKS)
344 		return -EINVAL;
345 
346 	for (i = 0; i < xcp_mgr->num_xcps; ++i) {
347 		xcp = &xcp_mgr->xcp[i];
348 		if ((xcp->valid) && (xcp->ip[ip].valid) &&
349 		    (xcp->ip[ip].inst_mask & BIT(instance)))
350 			id_mask |= BIT(i);
351 	}
352 
353 	if (!id_mask)
354 		id_mask = -ENXIO;
355 
356 	return id_mask;
357 }
358 
359 int amdgpu_xcp_get_inst_details(struct amdgpu_xcp *xcp,
360 				enum AMDGPU_XCP_IP_BLOCK ip,
361 				uint32_t *inst_mask)
362 {
363 	if (!xcp->valid || !inst_mask || !(xcp->ip[ip].valid))
364 		return -EINVAL;
365 
366 	*inst_mask = xcp->ip[ip].inst_mask;
367 
368 	return 0;
369 }
370 
371 int amdgpu_xcp_dev_register(struct amdgpu_device *adev,
372 			const struct pci_device_id *ent)
373 {
374 	int i, ret;
375 
376 	if (!adev->xcp_mgr)
377 		return 0;
378 
379 	for (i = 1; i < MAX_XCP; i++) {
380 		if (!adev->xcp_mgr->xcp[i].ddev)
381 			break;
382 
383 		ret = drm_dev_register(adev->xcp_mgr->xcp[i].ddev, ent->driver_data);
384 		if (ret)
385 			return ret;
386 	}
387 
388 	return 0;
389 }
390 
391 void amdgpu_xcp_dev_unplug(struct amdgpu_device *adev)
392 {
393 	struct drm_device *p_ddev;
394 	int i;
395 
396 	if (!adev->xcp_mgr)
397 		return;
398 
399 	for (i = 1; i < MAX_XCP; i++) {
400 		if (!adev->xcp_mgr->xcp[i].ddev)
401 			break;
402 
403 		p_ddev = adev->xcp_mgr->xcp[i].ddev;
404 		drm_dev_unplug(p_ddev);
405 		p_ddev->render->dev = adev->xcp_mgr->xcp[i].rdev;
406 		p_ddev->primary->dev = adev->xcp_mgr->xcp[i].pdev;
407 		p_ddev->driver =  adev->xcp_mgr->xcp[i].driver;
408 		p_ddev->vma_offset_manager = adev->xcp_mgr->xcp[i].vma_offset_manager;
409 		amdgpu_xcp_drm_dev_free(p_ddev);
410 	}
411 }
412 
413 int amdgpu_xcp_open_device(struct amdgpu_device *adev,
414 			   struct amdgpu_fpriv *fpriv,
415 			   struct drm_file *file_priv)
416 {
417 	int i;
418 
419 	if (!adev->xcp_mgr)
420 		return 0;
421 
422 	fpriv->xcp_id = AMDGPU_XCP_NO_PARTITION;
423 	for (i = 0; i < MAX_XCP; ++i) {
424 		if (!adev->xcp_mgr->xcp[i].ddev)
425 			break;
426 
427 		if (file_priv->minor == adev->xcp_mgr->xcp[i].ddev->render) {
428 			if (adev->xcp_mgr->xcp[i].valid == FALSE) {
429 				dev_err(adev->dev, "renderD%d partition %d not valid!",
430 						file_priv->minor->index, i);
431 				return -ENOENT;
432 			}
433 			dev_dbg(adev->dev, "renderD%d partition %d opened!",
434 					file_priv->minor->index, i);
435 			fpriv->xcp_id = i;
436 			break;
437 		}
438 	}
439 
440 	fpriv->vm.mem_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ? -1 :
441 				adev->xcp_mgr->xcp[fpriv->xcp_id].mem_id;
442 	return 0;
443 }
444 
445 void amdgpu_xcp_release_sched(struct amdgpu_device *adev,
446 				  struct amdgpu_ctx_entity *entity)
447 {
448 	struct drm_gpu_scheduler *sched;
449 	struct amdgpu_ring *ring;
450 
451 	if (!adev->xcp_mgr)
452 		return;
453 
454 	sched = entity->entity.rq->sched;
455 	if (drm_sched_wqueue_ready(sched)) {
456 		ring = to_amdgpu_ring(entity->entity.rq->sched);
457 		atomic_dec(&adev->xcp_mgr->xcp[ring->xcp_id].ref_cnt);
458 	}
459 }
460 
461 int amdgpu_xcp_select_scheds(struct amdgpu_device *adev,
462 			     u32 hw_ip, u32 hw_prio,
463 			     struct amdgpu_fpriv *fpriv,
464 			     unsigned int *num_scheds,
465 			     struct drm_gpu_scheduler ***scheds)
466 {
467 	u32 sel_xcp_id;
468 	int i;
469 	struct amdgpu_xcp_mgr *xcp_mgr = adev->xcp_mgr;
470 
471 	if (fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION) {
472 		u32 least_ref_cnt = ~0;
473 
474 		fpriv->xcp_id = 0;
475 		for (i = 0; i < xcp_mgr->num_xcps; i++) {
476 			u32 total_ref_cnt;
477 
478 			total_ref_cnt = atomic_read(&xcp_mgr->xcp[i].ref_cnt);
479 			if (total_ref_cnt < least_ref_cnt) {
480 				fpriv->xcp_id = i;
481 				least_ref_cnt = total_ref_cnt;
482 			}
483 		}
484 	}
485 	sel_xcp_id = fpriv->xcp_id;
486 
487 	if (xcp_mgr->xcp[sel_xcp_id].gpu_sched[hw_ip][hw_prio].num_scheds) {
488 		*num_scheds =
489 			xcp_mgr->xcp[fpriv->xcp_id].gpu_sched[hw_ip][hw_prio].num_scheds;
490 		*scheds =
491 			xcp_mgr->xcp[fpriv->xcp_id].gpu_sched[hw_ip][hw_prio].sched;
492 		atomic_inc(&adev->xcp_mgr->xcp[sel_xcp_id].ref_cnt);
493 		dev_dbg(adev->dev, "Selected partition #%d", sel_xcp_id);
494 	} else {
495 		dev_err(adev->dev, "Failed to schedule partition #%d.", sel_xcp_id);
496 		return -ENOENT;
497 	}
498 
499 	return 0;
500 }
501 
502 static void amdgpu_set_xcp_id(struct amdgpu_device *adev,
503 			      uint32_t inst_idx,
504 			      struct amdgpu_ring *ring)
505 {
506 	int xcp_id;
507 	enum AMDGPU_XCP_IP_BLOCK ip_blk;
508 	uint32_t inst_mask;
509 
510 	ring->xcp_id = AMDGPU_XCP_NO_PARTITION;
511 	if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
512 		adev->gfx.enforce_isolation[0].xcp_id = ring->xcp_id;
513 	if ((adev->xcp_mgr->mode == AMDGPU_XCP_MODE_NONE) ||
514 	    (ring->funcs->type == AMDGPU_RING_TYPE_CPER))
515 		return;
516 
517 	inst_mask = 1 << inst_idx;
518 
519 	switch (ring->funcs->type) {
520 	case AMDGPU_HW_IP_GFX:
521 	case AMDGPU_RING_TYPE_COMPUTE:
522 	case AMDGPU_RING_TYPE_KIQ:
523 		ip_blk = AMDGPU_XCP_GFX;
524 		break;
525 	case AMDGPU_RING_TYPE_SDMA:
526 		ip_blk = AMDGPU_XCP_SDMA;
527 		break;
528 	case AMDGPU_RING_TYPE_VCN_ENC:
529 	case AMDGPU_RING_TYPE_VCN_JPEG:
530 		ip_blk = AMDGPU_XCP_VCN;
531 		break;
532 	default:
533 		dev_err(adev->dev, "Not support ring type %d!", ring->funcs->type);
534 		return;
535 	}
536 
537 	for (xcp_id = 0; xcp_id < adev->xcp_mgr->num_xcps; xcp_id++) {
538 		if (adev->xcp_mgr->xcp[xcp_id].ip[ip_blk].inst_mask & inst_mask) {
539 			ring->xcp_id = xcp_id;
540 			dev_dbg(adev->dev, "ring:%s xcp_id :%u", ring->name,
541 				ring->xcp_id);
542 			if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE)
543 				adev->gfx.enforce_isolation[xcp_id].xcp_id = xcp_id;
544 			break;
545 		}
546 	}
547 }
548 
549 static void amdgpu_xcp_gpu_sched_update(struct amdgpu_device *adev,
550 					struct amdgpu_ring *ring,
551 					unsigned int sel_xcp_id)
552 {
553 	unsigned int *num_gpu_sched;
554 
555 	num_gpu_sched = &adev->xcp_mgr->xcp[sel_xcp_id]
556 			.gpu_sched[ring->funcs->type][ring->hw_prio].num_scheds;
557 	adev->xcp_mgr->xcp[sel_xcp_id].gpu_sched[ring->funcs->type][ring->hw_prio]
558 			.sched[(*num_gpu_sched)++] = &ring->sched;
559 	dev_dbg(adev->dev, "%s :[%d] gpu_sched[%d][%d] = %d",
560 		ring->name, sel_xcp_id, ring->funcs->type,
561 		ring->hw_prio, *num_gpu_sched);
562 }
563 
564 static int amdgpu_xcp_sched_list_update(struct amdgpu_device *adev)
565 {
566 	struct amdgpu_ring *ring;
567 	int i;
568 
569 	for (i = 0; i < MAX_XCP; i++) {
570 		atomic_set(&adev->xcp_mgr->xcp[i].ref_cnt, 0);
571 		memset(adev->xcp_mgr->xcp[i].gpu_sched, 0, sizeof(adev->xcp_mgr->xcp->gpu_sched));
572 	}
573 
574 	if (adev->xcp_mgr->mode == AMDGPU_XCP_MODE_NONE)
575 		return 0;
576 
577 	for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
578 		ring = adev->rings[i];
579 		if (!ring || !ring->sched.ready || ring->no_scheduler)
580 			continue;
581 
582 		amdgpu_xcp_gpu_sched_update(adev, ring, ring->xcp_id);
583 
584 		/* VCN may be shared by two partitions under CPX MODE in certain
585 		 * configs.
586 		 */
587 		if ((ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC ||
588 		     ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG) &&
589 		    (adev->xcp_mgr->num_xcps > adev->vcn.num_vcn_inst))
590 			amdgpu_xcp_gpu_sched_update(adev, ring, ring->xcp_id + 1);
591 	}
592 
593 	return 0;
594 }
595 
596 int amdgpu_xcp_update_partition_sched_list(struct amdgpu_device *adev)
597 {
598 	int i;
599 
600 	for (i = 0; i < adev->num_rings; i++) {
601 		struct amdgpu_ring *ring = adev->rings[i];
602 
603 		if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE ||
604 			ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
605 			amdgpu_set_xcp_id(adev, ring->xcc_id, ring);
606 		else
607 			amdgpu_set_xcp_id(adev, ring->me, ring);
608 	}
609 
610 	return amdgpu_xcp_sched_list_update(adev);
611 }
612 
613 void amdgpu_xcp_update_supported_modes(struct amdgpu_xcp_mgr *xcp_mgr)
614 {
615 	struct amdgpu_device *adev = xcp_mgr->adev;
616 
617 	xcp_mgr->supp_xcp_modes = 0;
618 
619 	switch (NUM_XCC(adev->gfx.xcc_mask)) {
620 	case 8:
621 		xcp_mgr->supp_xcp_modes = BIT(AMDGPU_SPX_PARTITION_MODE) |
622 					  BIT(AMDGPU_DPX_PARTITION_MODE) |
623 					  BIT(AMDGPU_QPX_PARTITION_MODE) |
624 					  BIT(AMDGPU_CPX_PARTITION_MODE);
625 		break;
626 	case 6:
627 		xcp_mgr->supp_xcp_modes = BIT(AMDGPU_SPX_PARTITION_MODE) |
628 					  BIT(AMDGPU_TPX_PARTITION_MODE) |
629 					  BIT(AMDGPU_CPX_PARTITION_MODE);
630 		break;
631 	case 4:
632 		xcp_mgr->supp_xcp_modes = BIT(AMDGPU_SPX_PARTITION_MODE) |
633 					  BIT(AMDGPU_DPX_PARTITION_MODE) |
634 					  BIT(AMDGPU_CPX_PARTITION_MODE);
635 		break;
636 	case 2:
637 		xcp_mgr->supp_xcp_modes = BIT(AMDGPU_SPX_PARTITION_MODE) |
638 					  BIT(AMDGPU_CPX_PARTITION_MODE);
639 		break;
640 	case 1:
641 		xcp_mgr->supp_xcp_modes = BIT(AMDGPU_SPX_PARTITION_MODE) |
642 					  BIT(AMDGPU_CPX_PARTITION_MODE);
643 		break;
644 
645 	default:
646 		break;
647 	}
648 }
649 
650 int amdgpu_xcp_pre_partition_switch(struct amdgpu_xcp_mgr *xcp_mgr, u32 flags)
651 {
652 	/* TODO:
653 	 * Stop user queues and threads, and make sure GPU is empty of work.
654 	 */
655 
656 	if (flags & AMDGPU_XCP_OPS_KFD)
657 		amdgpu_amdkfd_device_fini_sw(xcp_mgr->adev);
658 
659 	return 0;
660 }
661 
662 int amdgpu_xcp_post_partition_switch(struct amdgpu_xcp_mgr *xcp_mgr, u32 flags)
663 {
664 	int ret = 0;
665 
666 	if (flags & AMDGPU_XCP_OPS_KFD) {
667 		amdgpu_amdkfd_device_probe(xcp_mgr->adev);
668 		amdgpu_amdkfd_device_init(xcp_mgr->adev);
669 		/* If KFD init failed, return failure */
670 		if (!xcp_mgr->adev->kfd.init_complete)
671 			ret = -EIO;
672 	}
673 
674 	return ret;
675 }
676 
677 /*====================== xcp sysfs - configuration ======================*/
678 #define XCP_CFG_SYSFS_RES_ATTR_SHOW(_name)                         \
679 	static ssize_t amdgpu_xcp_res_sysfs_##_name##_show(        \
680 		struct amdgpu_xcp_res_details *xcp_res, char *buf) \
681 	{                                                          \
682 		return sysfs_emit(buf, "%d\n", xcp_res->_name);    \
683 	}
684 
685 struct amdgpu_xcp_res_sysfs_attribute {
686 	struct attribute attr;
687 	ssize_t (*show)(struct amdgpu_xcp_res_details *xcp_res, char *buf);
688 };
689 
690 #define XCP_CFG_SYSFS_RES_ATTR(_name)                                        \
691 	struct amdgpu_xcp_res_sysfs_attribute xcp_res_sysfs_attr_##_name = { \
692 		.attr = { .name = __stringify(_name), .mode = 0400 },        \
693 		.show = amdgpu_xcp_res_sysfs_##_name##_show,                 \
694 	}
695 
696 XCP_CFG_SYSFS_RES_ATTR_SHOW(num_inst)
697 XCP_CFG_SYSFS_RES_ATTR(num_inst);
698 XCP_CFG_SYSFS_RES_ATTR_SHOW(num_shared)
699 XCP_CFG_SYSFS_RES_ATTR(num_shared);
700 
701 #define XCP_CFG_SYSFS_RES_ATTR_PTR(_name) xcp_res_sysfs_attr_##_name.attr
702 
703 static struct attribute *xcp_cfg_res_sysfs_attrs[] = {
704 	&XCP_CFG_SYSFS_RES_ATTR_PTR(num_inst),
705 	&XCP_CFG_SYSFS_RES_ATTR_PTR(num_shared), NULL
706 };
707 
708 static const char *xcp_desc[] = {
709 	[AMDGPU_SPX_PARTITION_MODE] = "SPX",
710 	[AMDGPU_DPX_PARTITION_MODE] = "DPX",
711 	[AMDGPU_TPX_PARTITION_MODE] = "TPX",
712 	[AMDGPU_QPX_PARTITION_MODE] = "QPX",
713 	[AMDGPU_CPX_PARTITION_MODE] = "CPX",
714 };
715 
716 static const char *nps_desc[] = {
717 	[UNKNOWN_MEMORY_PARTITION_MODE] = "UNKNOWN",
718 	[AMDGPU_NPS1_PARTITION_MODE] = "NPS1",
719 	[AMDGPU_NPS2_PARTITION_MODE] = "NPS2",
720 	[AMDGPU_NPS3_PARTITION_MODE] = "NPS3",
721 	[AMDGPU_NPS4_PARTITION_MODE] = "NPS4",
722 	[AMDGPU_NPS6_PARTITION_MODE] = "NPS6",
723 	[AMDGPU_NPS8_PARTITION_MODE] = "NPS8",
724 };
725 
726 ATTRIBUTE_GROUPS(xcp_cfg_res_sysfs);
727 
728 #define to_xcp_attr(x) \
729 	container_of(x, struct amdgpu_xcp_res_sysfs_attribute, attr)
730 #define to_xcp_res(x) container_of(x, struct amdgpu_xcp_res_details, kobj)
731 
732 static ssize_t xcp_cfg_res_sysfs_attr_show(struct kobject *kobj,
733 					   struct attribute *attr, char *buf)
734 {
735 	struct amdgpu_xcp_res_sysfs_attribute *attribute;
736 	struct amdgpu_xcp_res_details *xcp_res;
737 
738 	attribute = to_xcp_attr(attr);
739 	xcp_res = to_xcp_res(kobj);
740 
741 	if (!attribute->show)
742 		return -EIO;
743 
744 	return attribute->show(xcp_res, buf);
745 }
746 
747 static const struct sysfs_ops xcp_cfg_res_sysfs_ops = {
748 	.show = xcp_cfg_res_sysfs_attr_show,
749 };
750 
751 static const struct kobj_type xcp_cfg_res_sysfs_ktype = {
752 	.sysfs_ops = &xcp_cfg_res_sysfs_ops,
753 	.default_groups = xcp_cfg_res_sysfs_groups,
754 };
755 
756 const char *xcp_res_names[] = {
757 	[AMDGPU_XCP_RES_XCC] = "xcc",
758 	[AMDGPU_XCP_RES_DMA] = "dma",
759 	[AMDGPU_XCP_RES_DEC] = "dec",
760 	[AMDGPU_XCP_RES_JPEG] = "jpeg",
761 };
762 
763 static int amdgpu_xcp_get_res_info(struct amdgpu_xcp_mgr *xcp_mgr,
764 				   int mode,
765 				   struct amdgpu_xcp_cfg *xcp_cfg)
766 {
767 	if (xcp_mgr->funcs && xcp_mgr->funcs->get_xcp_res_info)
768 		return xcp_mgr->funcs->get_xcp_res_info(xcp_mgr, mode, xcp_cfg);
769 
770 	return -EOPNOTSUPP;
771 }
772 
773 #define to_xcp_cfg(x) container_of(x, struct amdgpu_xcp_cfg, kobj)
774 static ssize_t supported_xcp_configs_show(struct kobject *kobj,
775 					  struct kobj_attribute *attr, char *buf)
776 {
777 	struct amdgpu_xcp_cfg *xcp_cfg = to_xcp_cfg(kobj);
778 	struct amdgpu_xcp_mgr *xcp_mgr = xcp_cfg->xcp_mgr;
779 	int size = 0, mode;
780 	char *sep = "";
781 
782 	if (!xcp_mgr || !xcp_mgr->supp_xcp_modes)
783 		return sysfs_emit(buf, "Not supported\n");
784 
785 	for_each_inst(mode, xcp_mgr->supp_xcp_modes) {
786 		size += sysfs_emit_at(buf, size, "%s%s", sep, xcp_desc[mode]);
787 		sep = ", ";
788 	}
789 
790 	size += sysfs_emit_at(buf, size, "\n");
791 
792 	return size;
793 }
794 
795 static ssize_t supported_nps_configs_show(struct kobject *kobj,
796 					  struct kobj_attribute *attr, char *buf)
797 {
798 	struct amdgpu_xcp_cfg *xcp_cfg = to_xcp_cfg(kobj);
799 	int size = 0, mode;
800 	char *sep = "";
801 
802 	if (!xcp_cfg || !xcp_cfg->compatible_nps_modes)
803 		return sysfs_emit(buf, "Not supported\n");
804 
805 	for_each_inst(mode, xcp_cfg->compatible_nps_modes) {
806 		size += sysfs_emit_at(buf, size, "%s%s", sep, nps_desc[mode]);
807 		sep = ", ";
808 	}
809 
810 	size += sysfs_emit_at(buf, size, "\n");
811 
812 	return size;
813 }
814 
815 static ssize_t xcp_config_show(struct kobject *kobj,
816 			       struct kobj_attribute *attr, char *buf)
817 {
818 	struct amdgpu_xcp_cfg *xcp_cfg = to_xcp_cfg(kobj);
819 
820 	return sysfs_emit(buf, "%s\n",
821 			  amdgpu_gfx_compute_mode_desc(xcp_cfg->mode));
822 }
823 
824 static ssize_t xcp_config_store(struct kobject *kobj,
825 				struct kobj_attribute *attr,
826 				const char *buf, size_t size)
827 {
828 	struct amdgpu_xcp_cfg *xcp_cfg = to_xcp_cfg(kobj);
829 	int mode, r;
830 
831 	if (!strncasecmp("SPX", buf, strlen("SPX")))
832 		mode = AMDGPU_SPX_PARTITION_MODE;
833 	else if (!strncasecmp("DPX", buf, strlen("DPX")))
834 		mode = AMDGPU_DPX_PARTITION_MODE;
835 	else if (!strncasecmp("TPX", buf, strlen("TPX")))
836 		mode = AMDGPU_TPX_PARTITION_MODE;
837 	else if (!strncasecmp("QPX", buf, strlen("QPX")))
838 		mode = AMDGPU_QPX_PARTITION_MODE;
839 	else if (!strncasecmp("CPX", buf, strlen("CPX")))
840 		mode = AMDGPU_CPX_PARTITION_MODE;
841 	else
842 		return -EINVAL;
843 
844 	r = amdgpu_xcp_get_res_info(xcp_cfg->xcp_mgr, mode, xcp_cfg);
845 
846 	if (r)
847 		return r;
848 
849 	xcp_cfg->mode = mode;
850 	return size;
851 }
852 
853 static struct kobj_attribute xcp_cfg_sysfs_mode =
854 	__ATTR_RW_MODE(xcp_config, 0644);
855 
856 static void xcp_cfg_sysfs_release(struct kobject *kobj)
857 {
858 	struct amdgpu_xcp_cfg *xcp_cfg = to_xcp_cfg(kobj);
859 
860 	kfree(xcp_cfg);
861 }
862 
863 static const struct kobj_type xcp_cfg_sysfs_ktype = {
864 	.release = xcp_cfg_sysfs_release,
865 	.sysfs_ops = &kobj_sysfs_ops,
866 };
867 
868 static struct kobj_attribute supp_part_sysfs_mode =
869 	__ATTR_RO(supported_xcp_configs);
870 
871 static struct kobj_attribute supp_nps_sysfs_mode =
872 	__ATTR_RO(supported_nps_configs);
873 
874 static const struct attribute *xcp_attrs[] = {
875 	&supp_part_sysfs_mode.attr,
876 	&xcp_cfg_sysfs_mode.attr,
877 	NULL,
878 };
879 
880 static void amdgpu_xcp_cfg_sysfs_init(struct amdgpu_device *adev)
881 {
882 	struct amdgpu_xcp_res_details *xcp_res;
883 	struct amdgpu_xcp_cfg *xcp_cfg;
884 	int i, r, j, rid, mode;
885 
886 	if (!adev->xcp_mgr)
887 		return;
888 
889 	xcp_cfg = kzalloc(sizeof(*xcp_cfg), GFP_KERNEL);
890 	if (!xcp_cfg)
891 		return;
892 	xcp_cfg->xcp_mgr = adev->xcp_mgr;
893 
894 	r = kobject_init_and_add(&xcp_cfg->kobj, &xcp_cfg_sysfs_ktype,
895 				 &adev->dev->kobj, "compute_partition_config");
896 	if (r)
897 		goto err1;
898 
899 	r = sysfs_create_files(&xcp_cfg->kobj, xcp_attrs);
900 	if (r)
901 		goto err1;
902 
903 	if (adev->gmc.supported_nps_modes != 0) {
904 		r = sysfs_create_file(&xcp_cfg->kobj, &supp_nps_sysfs_mode.attr);
905 		if (r) {
906 			sysfs_remove_files(&xcp_cfg->kobj, xcp_attrs);
907 			goto err1;
908 		}
909 	}
910 
911 	mode = (xcp_cfg->xcp_mgr->mode ==
912 		AMDGPU_UNKNOWN_COMPUTE_PARTITION_MODE) ?
913 		       AMDGPU_SPX_PARTITION_MODE :
914 		       xcp_cfg->xcp_mgr->mode;
915 	r = amdgpu_xcp_get_res_info(xcp_cfg->xcp_mgr, mode, xcp_cfg);
916 	if (r) {
917 		sysfs_remove_file(&xcp_cfg->kobj, &supp_nps_sysfs_mode.attr);
918 		sysfs_remove_files(&xcp_cfg->kobj, xcp_attrs);
919 		goto err1;
920 	}
921 
922 	xcp_cfg->mode = mode;
923 	for (i = 0; i < xcp_cfg->num_res; i++) {
924 		xcp_res = &xcp_cfg->xcp_res[i];
925 		rid = xcp_res->id;
926 		r = kobject_init_and_add(&xcp_res->kobj,
927 					 &xcp_cfg_res_sysfs_ktype,
928 					 &xcp_cfg->kobj, "%s",
929 					 xcp_res_names[rid]);
930 		if (r)
931 			goto err;
932 	}
933 
934 	adev->xcp_mgr->xcp_cfg = xcp_cfg;
935 	return;
936 err:
937 	for (j = 0; j < i; j++) {
938 		xcp_res = &xcp_cfg->xcp_res[i];
939 		kobject_put(&xcp_res->kobj);
940 	}
941 
942 	sysfs_remove_file(&xcp_cfg->kobj, &supp_nps_sysfs_mode.attr);
943 	sysfs_remove_files(&xcp_cfg->kobj, xcp_attrs);
944 err1:
945 	kobject_put(&xcp_cfg->kobj);
946 }
947 
948 static void amdgpu_xcp_cfg_sysfs_fini(struct amdgpu_device *adev)
949 {
950 	struct amdgpu_xcp_res_details *xcp_res;
951 	struct amdgpu_xcp_cfg *xcp_cfg;
952 	int i;
953 
954 	if (!adev->xcp_mgr || !adev->xcp_mgr->xcp_cfg)
955 		return;
956 
957 	xcp_cfg = adev->xcp_mgr->xcp_cfg;
958 	for (i = 0; i < xcp_cfg->num_res; i++) {
959 		xcp_res = &xcp_cfg->xcp_res[i];
960 		kobject_put(&xcp_res->kobj);
961 	}
962 
963 	sysfs_remove_file(&xcp_cfg->kobj, &supp_nps_sysfs_mode.attr);
964 	sysfs_remove_files(&xcp_cfg->kobj, xcp_attrs);
965 	kobject_put(&xcp_cfg->kobj);
966 }
967 
968 /*====================== xcp sysfs - data entries ======================*/
969 
970 #define to_xcp(x) container_of(x, struct amdgpu_xcp, kobj)
971 
972 static ssize_t xcp_metrics_show(struct kobject *kobj,
973 				struct kobj_attribute *attr, char *buf)
974 {
975 	struct amdgpu_xcp *xcp = to_xcp(kobj);
976 	struct amdgpu_xcp_mgr *xcp_mgr;
977 	ssize_t size;
978 
979 	xcp_mgr = xcp->xcp_mgr;
980 	size = amdgpu_dpm_get_xcp_metrics(xcp_mgr->adev, xcp->id, NULL);
981 	if (size <= 0)
982 		return size;
983 
984 	if (size > PAGE_SIZE)
985 		return -ENOSPC;
986 
987 	return amdgpu_dpm_get_xcp_metrics(xcp_mgr->adev, xcp->id, buf);
988 }
989 
990 static umode_t amdgpu_xcp_attrs_is_visible(struct kobject *kobj,
991 					   struct attribute *attr, int n)
992 {
993 	struct amdgpu_xcp *xcp = to_xcp(kobj);
994 
995 	if (!xcp || !xcp->valid)
996 		return 0;
997 
998 	return attr->mode;
999 }
1000 
1001 static struct kobj_attribute xcp_sysfs_metrics = __ATTR_RO(xcp_metrics);
1002 
1003 static struct attribute *amdgpu_xcp_attrs[] = {
1004 	&xcp_sysfs_metrics.attr,
1005 	NULL,
1006 };
1007 
1008 static const struct attribute_group amdgpu_xcp_attrs_group = {
1009 	.attrs = amdgpu_xcp_attrs,
1010 	.is_visible = amdgpu_xcp_attrs_is_visible
1011 };
1012 
1013 static const struct kobj_type xcp_sysfs_ktype = {
1014 	.sysfs_ops = &kobj_sysfs_ops,
1015 };
1016 
1017 static void amdgpu_xcp_sysfs_entries_fini(struct amdgpu_xcp_mgr *xcp_mgr, int n)
1018 {
1019 	struct amdgpu_xcp *xcp;
1020 
1021 	for (n--; n >= 0; n--) {
1022 		xcp = &xcp_mgr->xcp[n];
1023 		if (!xcp->ddev || !xcp->valid)
1024 			continue;
1025 		sysfs_remove_group(&xcp->kobj, &amdgpu_xcp_attrs_group);
1026 		kobject_put(&xcp->kobj);
1027 	}
1028 }
1029 
1030 static void amdgpu_xcp_sysfs_entries_init(struct amdgpu_xcp_mgr *xcp_mgr)
1031 {
1032 	struct amdgpu_xcp *xcp;
1033 	int i, r;
1034 
1035 	for (i = 0; i < MAX_XCP; i++) {
1036 		/* Redirect all IOCTLs to the primary device */
1037 		xcp = &xcp_mgr->xcp[i];
1038 		if (!xcp->ddev)
1039 			break;
1040 		r = kobject_init_and_add(&xcp->kobj, &xcp_sysfs_ktype,
1041 					 &xcp->ddev->dev->kobj, "xcp");
1042 		if (r)
1043 			goto out;
1044 
1045 		r = sysfs_create_group(&xcp->kobj, &amdgpu_xcp_attrs_group);
1046 		if (r)
1047 			goto out;
1048 	}
1049 
1050 	return;
1051 out:
1052 	kobject_put(&xcp->kobj);
1053 }
1054 
1055 static void amdgpu_xcp_sysfs_entries_update(struct amdgpu_xcp_mgr *xcp_mgr)
1056 {
1057 	struct amdgpu_xcp *xcp;
1058 	int i;
1059 
1060 	for (i = 0; i < MAX_XCP; i++) {
1061 		/* Redirect all IOCTLs to the primary device */
1062 		xcp = &xcp_mgr->xcp[i];
1063 		if (!xcp->ddev)
1064 			continue;
1065 		sysfs_update_group(&xcp->kobj, &amdgpu_xcp_attrs_group);
1066 	}
1067 
1068 	return;
1069 }
1070 
1071 void amdgpu_xcp_sysfs_init(struct amdgpu_device *adev)
1072 {
1073 	if (!adev->xcp_mgr)
1074 		return;
1075 
1076 	amdgpu_xcp_cfg_sysfs_init(adev);
1077 
1078 	return;
1079 }
1080 
1081 void amdgpu_xcp_sysfs_fini(struct amdgpu_device *adev)
1082 {
1083 	if (!adev->xcp_mgr)
1084 		return;
1085 	amdgpu_xcp_sysfs_entries_fini(adev->xcp_mgr, MAX_XCP);
1086 	amdgpu_xcp_cfg_sysfs_fini(adev);
1087 }
1088