xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c (revision a1ff5a7d78a036d6c2178ee5acd6ba4946243800)
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 
__amdgpu_xcp_run(struct amdgpu_xcp_mgr * xcp_mgr,struct amdgpu_xcp_ip * xcp_ip,int xcp_state)30 static int __amdgpu_xcp_run(struct amdgpu_xcp_mgr *xcp_mgr,
31 			    struct amdgpu_xcp_ip *xcp_ip, int xcp_state)
32 {
33 	int (*run_func)(void *handle, uint32_t inst_mask);
34 	int ret = 0;
35 
36 	if (!xcp_ip || !xcp_ip->valid || !xcp_ip->ip_funcs)
37 		return 0;
38 
39 	run_func = NULL;
40 
41 	switch (xcp_state) {
42 	case AMDGPU_XCP_PREPARE_SUSPEND:
43 		run_func = xcp_ip->ip_funcs->prepare_suspend;
44 		break;
45 	case AMDGPU_XCP_SUSPEND:
46 		run_func = xcp_ip->ip_funcs->suspend;
47 		break;
48 	case AMDGPU_XCP_PREPARE_RESUME:
49 		run_func = xcp_ip->ip_funcs->prepare_resume;
50 		break;
51 	case AMDGPU_XCP_RESUME:
52 		run_func = xcp_ip->ip_funcs->resume;
53 		break;
54 	}
55 
56 	if (run_func)
57 		ret = run_func(xcp_mgr->adev, xcp_ip->inst_mask);
58 
59 	return ret;
60 }
61 
amdgpu_xcp_run_transition(struct amdgpu_xcp_mgr * xcp_mgr,int xcp_id,int state)62 static int amdgpu_xcp_run_transition(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id,
63 				     int state)
64 {
65 	struct amdgpu_xcp_ip *xcp_ip;
66 	struct amdgpu_xcp *xcp;
67 	int i, ret;
68 
69 	if (xcp_id >= MAX_XCP || !xcp_mgr->xcp[xcp_id].valid)
70 		return -EINVAL;
71 
72 	xcp = &xcp_mgr->xcp[xcp_id];
73 	for (i = 0; i < AMDGPU_XCP_MAX_BLOCKS; ++i) {
74 		xcp_ip = &xcp->ip[i];
75 		ret = __amdgpu_xcp_run(xcp_mgr, xcp_ip, state);
76 		if (ret)
77 			break;
78 	}
79 
80 	return ret;
81 }
82 
amdgpu_xcp_prepare_suspend(struct amdgpu_xcp_mgr * xcp_mgr,int xcp_id)83 int amdgpu_xcp_prepare_suspend(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id)
84 {
85 	return amdgpu_xcp_run_transition(xcp_mgr, xcp_id,
86 					 AMDGPU_XCP_PREPARE_SUSPEND);
87 }
88 
amdgpu_xcp_suspend(struct amdgpu_xcp_mgr * xcp_mgr,int xcp_id)89 int amdgpu_xcp_suspend(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id)
90 {
91 	return amdgpu_xcp_run_transition(xcp_mgr, xcp_id, AMDGPU_XCP_SUSPEND);
92 }
93 
amdgpu_xcp_prepare_resume(struct amdgpu_xcp_mgr * xcp_mgr,int xcp_id)94 int amdgpu_xcp_prepare_resume(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id)
95 {
96 	return amdgpu_xcp_run_transition(xcp_mgr, xcp_id,
97 					 AMDGPU_XCP_PREPARE_RESUME);
98 }
99 
amdgpu_xcp_resume(struct amdgpu_xcp_mgr * xcp_mgr,int xcp_id)100 int amdgpu_xcp_resume(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id)
101 {
102 	return amdgpu_xcp_run_transition(xcp_mgr, xcp_id, AMDGPU_XCP_RESUME);
103 }
104 
__amdgpu_xcp_add_block(struct amdgpu_xcp_mgr * xcp_mgr,int xcp_id,struct amdgpu_xcp_ip * ip)105 static void __amdgpu_xcp_add_block(struct amdgpu_xcp_mgr *xcp_mgr, int xcp_id,
106 				   struct amdgpu_xcp_ip *ip)
107 {
108 	struct amdgpu_xcp *xcp;
109 
110 	if (!ip)
111 		return;
112 
113 	xcp = &xcp_mgr->xcp[xcp_id];
114 	xcp->ip[ip->ip_id] = *ip;
115 	xcp->ip[ip->ip_id].valid = true;
116 
117 	xcp->valid = true;
118 }
119 
amdgpu_xcp_init(struct amdgpu_xcp_mgr * xcp_mgr,int num_xcps,int mode)120 int amdgpu_xcp_init(struct amdgpu_xcp_mgr *xcp_mgr, int num_xcps, int mode)
121 {
122 	struct amdgpu_device *adev = xcp_mgr->adev;
123 	struct amdgpu_xcp_ip ip;
124 	uint8_t mem_id;
125 	int i, j, ret;
126 
127 	if (!num_xcps || num_xcps > MAX_XCP)
128 		return -EINVAL;
129 
130 	xcp_mgr->mode = mode;
131 
132 	for (i = 0; i < MAX_XCP; ++i)
133 		xcp_mgr->xcp[i].valid = false;
134 
135 	/* This is needed for figuring out memory id of xcp */
136 	xcp_mgr->num_xcp_per_mem_partition = num_xcps / xcp_mgr->adev->gmc.num_mem_partitions;
137 
138 	for (i = 0; i < num_xcps; ++i) {
139 		for (j = AMDGPU_XCP_GFXHUB; j < AMDGPU_XCP_MAX_BLOCKS; ++j) {
140 			ret = xcp_mgr->funcs->get_ip_details(xcp_mgr, i, j,
141 							     &ip);
142 			if (ret)
143 				continue;
144 
145 			__amdgpu_xcp_add_block(xcp_mgr, i, &ip);
146 		}
147 
148 		xcp_mgr->xcp[i].id = i;
149 
150 		if (xcp_mgr->funcs->get_xcp_mem_id) {
151 			ret = xcp_mgr->funcs->get_xcp_mem_id(
152 				xcp_mgr, &xcp_mgr->xcp[i], &mem_id);
153 			if (ret)
154 				continue;
155 			else
156 				xcp_mgr->xcp[i].mem_id = mem_id;
157 		}
158 	}
159 
160 	xcp_mgr->num_xcps = num_xcps;
161 	amdgpu_xcp_update_partition_sched_list(adev);
162 
163 	return 0;
164 }
165 
__amdgpu_xcp_switch_partition_mode(struct amdgpu_xcp_mgr * xcp_mgr,int mode)166 static int __amdgpu_xcp_switch_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr,
167 					      int mode)
168 {
169 	int ret, curr_mode, num_xcps = 0;
170 
171 	if (!xcp_mgr->funcs || !xcp_mgr->funcs->switch_partition_mode)
172 		return 0;
173 
174 	mutex_lock(&xcp_mgr->xcp_lock);
175 
176 	curr_mode = xcp_mgr->mode;
177 	/* State set to transient mode */
178 	xcp_mgr->mode = AMDGPU_XCP_MODE_TRANS;
179 
180 	ret = xcp_mgr->funcs->switch_partition_mode(xcp_mgr, mode, &num_xcps);
181 
182 	if (ret) {
183 		/* Failed, get whatever mode it's at now */
184 		if (xcp_mgr->funcs->query_partition_mode)
185 			xcp_mgr->mode = amdgpu_xcp_query_partition_mode(
186 				xcp_mgr, AMDGPU_XCP_FL_LOCKED);
187 		else
188 			xcp_mgr->mode = curr_mode;
189 
190 		goto out;
191 	}
192 
193 out:
194 	mutex_unlock(&xcp_mgr->xcp_lock);
195 
196 	return ret;
197 }
198 
amdgpu_xcp_switch_partition_mode(struct amdgpu_xcp_mgr * xcp_mgr,int mode)199 int amdgpu_xcp_switch_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, int mode)
200 {
201 	if (!xcp_mgr || mode == AMDGPU_XCP_MODE_NONE)
202 		return -EINVAL;
203 
204 	if (xcp_mgr->mode == mode)
205 		return 0;
206 
207 	return __amdgpu_xcp_switch_partition_mode(xcp_mgr, mode);
208 }
209 
amdgpu_xcp_restore_partition_mode(struct amdgpu_xcp_mgr * xcp_mgr)210 int amdgpu_xcp_restore_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr)
211 {
212 	if (!xcp_mgr || xcp_mgr->mode == AMDGPU_XCP_MODE_NONE)
213 		return 0;
214 
215 	return __amdgpu_xcp_switch_partition_mode(xcp_mgr, xcp_mgr->mode);
216 }
217 
amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr * xcp_mgr,u32 flags)218 int amdgpu_xcp_query_partition_mode(struct amdgpu_xcp_mgr *xcp_mgr, u32 flags)
219 {
220 	int mode;
221 
222 	if (!amdgpu_sriov_vf(xcp_mgr->adev) &&
223 	    xcp_mgr->mode == AMDGPU_XCP_MODE_NONE)
224 		return xcp_mgr->mode;
225 
226 	if (!xcp_mgr->funcs || !xcp_mgr->funcs->query_partition_mode)
227 		return xcp_mgr->mode;
228 
229 	if (!(flags & AMDGPU_XCP_FL_LOCKED))
230 		mutex_lock(&xcp_mgr->xcp_lock);
231 	mode = xcp_mgr->funcs->query_partition_mode(xcp_mgr);
232 
233 	/* First time query for VF, set the mode here */
234 	if (amdgpu_sriov_vf(xcp_mgr->adev) &&
235 	    xcp_mgr->mode == AMDGPU_XCP_MODE_NONE)
236 		xcp_mgr->mode = mode;
237 
238 	if (xcp_mgr->mode != AMDGPU_XCP_MODE_TRANS && mode != xcp_mgr->mode)
239 		dev_WARN(
240 			xcp_mgr->adev->dev,
241 			"Cached partition mode %d not matching with device mode %d",
242 			xcp_mgr->mode, mode);
243 
244 	if (!(flags & AMDGPU_XCP_FL_LOCKED))
245 		mutex_unlock(&xcp_mgr->xcp_lock);
246 
247 	return mode;
248 }
249 
amdgpu_xcp_dev_alloc(struct amdgpu_device * adev)250 static int amdgpu_xcp_dev_alloc(struct amdgpu_device *adev)
251 {
252 	struct drm_device *p_ddev;
253 	struct drm_device *ddev;
254 	int i, ret;
255 
256 	ddev = adev_to_drm(adev);
257 
258 	/* xcp #0 shares drm device setting with adev */
259 	adev->xcp_mgr->xcp->ddev = ddev;
260 
261 	for (i = 1; i < MAX_XCP; i++) {
262 		ret = amdgpu_xcp_drm_dev_alloc(&p_ddev);
263 		if (ret == -ENOSPC) {
264 			dev_warn(adev->dev,
265 			"Skip xcp node #%d when out of drm node resource.", i);
266 			return 0;
267 		} else if (ret) {
268 			return ret;
269 		}
270 
271 		/* Redirect all IOCTLs to the primary device */
272 		adev->xcp_mgr->xcp[i].rdev = p_ddev->render->dev;
273 		adev->xcp_mgr->xcp[i].pdev = p_ddev->primary->dev;
274 		adev->xcp_mgr->xcp[i].driver = (struct drm_driver *)p_ddev->driver;
275 		adev->xcp_mgr->xcp[i].vma_offset_manager = p_ddev->vma_offset_manager;
276 		p_ddev->render->dev = ddev;
277 		p_ddev->primary->dev = ddev;
278 		p_ddev->vma_offset_manager = ddev->vma_offset_manager;
279 		p_ddev->driver = &amdgpu_partition_driver;
280 		adev->xcp_mgr->xcp[i].ddev = p_ddev;
281 	}
282 
283 	return 0;
284 }
285 
amdgpu_xcp_mgr_init(struct amdgpu_device * adev,int init_mode,int init_num_xcps,struct amdgpu_xcp_mgr_funcs * xcp_funcs)286 int amdgpu_xcp_mgr_init(struct amdgpu_device *adev, int init_mode,
287 			int init_num_xcps,
288 			struct amdgpu_xcp_mgr_funcs *xcp_funcs)
289 {
290 	struct amdgpu_xcp_mgr *xcp_mgr;
291 
292 	if (!xcp_funcs || !xcp_funcs->get_ip_details)
293 		return -EINVAL;
294 
295 	xcp_mgr = kzalloc(sizeof(*xcp_mgr), GFP_KERNEL);
296 
297 	if (!xcp_mgr)
298 		return -ENOMEM;
299 
300 	xcp_mgr->adev = adev;
301 	xcp_mgr->funcs = xcp_funcs;
302 	xcp_mgr->mode = init_mode;
303 	mutex_init(&xcp_mgr->xcp_lock);
304 
305 	if (init_mode != AMDGPU_XCP_MODE_NONE)
306 		amdgpu_xcp_init(xcp_mgr, init_num_xcps, init_mode);
307 
308 	adev->xcp_mgr = xcp_mgr;
309 
310 	return amdgpu_xcp_dev_alloc(adev);
311 }
312 
amdgpu_xcp_get_partition(struct amdgpu_xcp_mgr * xcp_mgr,enum AMDGPU_XCP_IP_BLOCK ip,int instance)313 int amdgpu_xcp_get_partition(struct amdgpu_xcp_mgr *xcp_mgr,
314 			     enum AMDGPU_XCP_IP_BLOCK ip, int instance)
315 {
316 	struct amdgpu_xcp *xcp;
317 	int i, id_mask = 0;
318 
319 	if (ip >= AMDGPU_XCP_MAX_BLOCKS)
320 		return -EINVAL;
321 
322 	for (i = 0; i < xcp_mgr->num_xcps; ++i) {
323 		xcp = &xcp_mgr->xcp[i];
324 		if ((xcp->valid) && (xcp->ip[ip].valid) &&
325 		    (xcp->ip[ip].inst_mask & BIT(instance)))
326 			id_mask |= BIT(i);
327 	}
328 
329 	if (!id_mask)
330 		id_mask = -ENXIO;
331 
332 	return id_mask;
333 }
334 
amdgpu_xcp_get_inst_details(struct amdgpu_xcp * xcp,enum AMDGPU_XCP_IP_BLOCK ip,uint32_t * inst_mask)335 int amdgpu_xcp_get_inst_details(struct amdgpu_xcp *xcp,
336 				enum AMDGPU_XCP_IP_BLOCK ip,
337 				uint32_t *inst_mask)
338 {
339 	if (!xcp->valid || !inst_mask || !(xcp->ip[ip].valid))
340 		return -EINVAL;
341 
342 	*inst_mask = xcp->ip[ip].inst_mask;
343 
344 	return 0;
345 }
346 
amdgpu_xcp_dev_register(struct amdgpu_device * adev,const struct pci_device_id * ent)347 int amdgpu_xcp_dev_register(struct amdgpu_device *adev,
348 			const struct pci_device_id *ent)
349 {
350 	int i, ret;
351 
352 	if (!adev->xcp_mgr)
353 		return 0;
354 
355 	for (i = 1; i < MAX_XCP; i++) {
356 		if (!adev->xcp_mgr->xcp[i].ddev)
357 			break;
358 
359 		ret = drm_dev_register(adev->xcp_mgr->xcp[i].ddev, ent->driver_data);
360 		if (ret)
361 			return ret;
362 	}
363 
364 	return 0;
365 }
366 
amdgpu_xcp_dev_unplug(struct amdgpu_device * adev)367 void amdgpu_xcp_dev_unplug(struct amdgpu_device *adev)
368 {
369 	struct drm_device *p_ddev;
370 	int i;
371 
372 	if (!adev->xcp_mgr)
373 		return;
374 
375 	for (i = 1; i < MAX_XCP; i++) {
376 		if (!adev->xcp_mgr->xcp[i].ddev)
377 			break;
378 
379 		p_ddev = adev->xcp_mgr->xcp[i].ddev;
380 		drm_dev_unplug(p_ddev);
381 		p_ddev->render->dev = adev->xcp_mgr->xcp[i].rdev;
382 		p_ddev->primary->dev = adev->xcp_mgr->xcp[i].pdev;
383 		p_ddev->driver =  adev->xcp_mgr->xcp[i].driver;
384 		p_ddev->vma_offset_manager = adev->xcp_mgr->xcp[i].vma_offset_manager;
385 	}
386 }
387 
amdgpu_xcp_open_device(struct amdgpu_device * adev,struct amdgpu_fpriv * fpriv,struct drm_file * file_priv)388 int amdgpu_xcp_open_device(struct amdgpu_device *adev,
389 			   struct amdgpu_fpriv *fpriv,
390 			   struct drm_file *file_priv)
391 {
392 	int i;
393 
394 	if (!adev->xcp_mgr)
395 		return 0;
396 
397 	fpriv->xcp_id = AMDGPU_XCP_NO_PARTITION;
398 	for (i = 0; i < MAX_XCP; ++i) {
399 		if (!adev->xcp_mgr->xcp[i].ddev)
400 			break;
401 
402 		if (file_priv->minor == adev->xcp_mgr->xcp[i].ddev->render) {
403 			if (adev->xcp_mgr->xcp[i].valid == FALSE) {
404 				dev_err(adev->dev, "renderD%d partition %d not valid!",
405 						file_priv->minor->index, i);
406 				return -ENOENT;
407 			}
408 			dev_dbg(adev->dev, "renderD%d partition %d opened!",
409 					file_priv->minor->index, i);
410 			fpriv->xcp_id = i;
411 			break;
412 		}
413 	}
414 
415 	fpriv->vm.mem_id = fpriv->xcp_id == AMDGPU_XCP_NO_PARTITION ? -1 :
416 				adev->xcp_mgr->xcp[fpriv->xcp_id].mem_id;
417 	return 0;
418 }
419 
amdgpu_xcp_release_sched(struct amdgpu_device * adev,struct amdgpu_ctx_entity * entity)420 void amdgpu_xcp_release_sched(struct amdgpu_device *adev,
421 				  struct amdgpu_ctx_entity *entity)
422 {
423 	struct drm_gpu_scheduler *sched;
424 	struct amdgpu_ring *ring;
425 
426 	if (!adev->xcp_mgr)
427 		return;
428 
429 	sched = entity->entity.rq->sched;
430 	if (sched->ready) {
431 		ring = to_amdgpu_ring(entity->entity.rq->sched);
432 		atomic_dec(&adev->xcp_mgr->xcp[ring->xcp_id].ref_cnt);
433 	}
434 }
435 
436