xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c (revision 4f727ecefefbd180de10e25b3e74c03dce3f1e75)
1 /*
2  * Copyright 2018 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  */
24 #include <linux/list.h>
25 #include "amdgpu.h"
26 #include "amdgpu_xgmi.h"
27 #include "amdgpu_smu.h"
28 
29 
30 static DEFINE_MUTEX(xgmi_mutex);
31 
32 #define AMDGPU_MAX_XGMI_HIVE			8
33 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE		4
34 
35 static struct amdgpu_hive_info xgmi_hives[AMDGPU_MAX_XGMI_HIVE];
36 static unsigned hive_count = 0;
37 
38 void *amdgpu_xgmi_hive_try_lock(struct amdgpu_hive_info *hive)
39 {
40 	return &hive->device_list;
41 }
42 
43 /**
44  * DOC: AMDGPU XGMI Support
45  *
46  * XGMI is a high speed interconnect that joins multiple GPU cards
47  * into a homogeneous memory space that is organized by a collective
48  * hive ID and individual node IDs, both of which are 64-bit numbers.
49  *
50  * The file xgmi_device_id contains the unique per GPU device ID and
51  * is stored in the /sys/class/drm/card${cardno}/device/ directory.
52  *
53  * Inside the device directory a sub-directory 'xgmi_hive_info' is
54  * created which contains the hive ID and the list of nodes.
55  *
56  * The hive ID is stored in:
57  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id
58  *
59  * The node information is stored in numbered directories:
60  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id
61  *
62  * Each device has their own xgmi_hive_info direction with a mirror
63  * set of node sub-directories.
64  *
65  * The XGMI memory space is built by contiguously adding the power of
66  * two padded VRAM space from each node to each other.
67  *
68  */
69 
70 
71 static ssize_t amdgpu_xgmi_show_hive_id(struct device *dev,
72 		struct device_attribute *attr, char *buf)
73 {
74 	struct amdgpu_hive_info *hive =
75 			container_of(attr, struct amdgpu_hive_info, dev_attr);
76 
77 	return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id);
78 }
79 
80 static int amdgpu_xgmi_sysfs_create(struct amdgpu_device *adev,
81 				    struct amdgpu_hive_info *hive)
82 {
83 	int ret = 0;
84 
85 	if (WARN_ON(hive->kobj))
86 		return -EINVAL;
87 
88 	hive->kobj = kobject_create_and_add("xgmi_hive_info", &adev->dev->kobj);
89 	if (!hive->kobj) {
90 		dev_err(adev->dev, "XGMI: Failed to allocate sysfs entry!\n");
91 		return -EINVAL;
92 	}
93 
94 	hive->dev_attr = (struct device_attribute) {
95 		.attr = {
96 			.name = "xgmi_hive_id",
97 			.mode = S_IRUGO,
98 
99 		},
100 		.show = amdgpu_xgmi_show_hive_id,
101 	};
102 
103 	ret = sysfs_create_file(hive->kobj, &hive->dev_attr.attr);
104 	if (ret) {
105 		dev_err(adev->dev, "XGMI: Failed to create device file xgmi_hive_id\n");
106 		kobject_del(hive->kobj);
107 		kobject_put(hive->kobj);
108 		hive->kobj = NULL;
109 	}
110 
111 	return ret;
112 }
113 
114 static void amdgpu_xgmi_sysfs_destroy(struct amdgpu_device *adev,
115 				    struct amdgpu_hive_info *hive)
116 {
117 	sysfs_remove_file(hive->kobj, &hive->dev_attr.attr);
118 	kobject_del(hive->kobj);
119 	kobject_put(hive->kobj);
120 	hive->kobj = NULL;
121 }
122 
123 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev,
124 				     struct device_attribute *attr,
125 				     char *buf)
126 {
127 	struct drm_device *ddev = dev_get_drvdata(dev);
128 	struct amdgpu_device *adev = ddev->dev_private;
129 
130 	return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.xgmi.node_id);
131 
132 }
133 
134 
135 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL);
136 
137 
138 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev,
139 					 struct amdgpu_hive_info *hive)
140 {
141 	int ret = 0;
142 	char node[10] = { 0 };
143 
144 	/* Create xgmi device id file */
145 	ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id);
146 	if (ret) {
147 		dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n");
148 		return ret;
149 	}
150 
151 	/* Create sysfs link to hive info folder on the first device */
152 	if (adev != hive->adev) {
153 		ret = sysfs_create_link(&adev->dev->kobj, hive->kobj,
154 					"xgmi_hive_info");
155 		if (ret) {
156 			dev_err(adev->dev, "XGMI: Failed to create link to hive info");
157 			goto remove_file;
158 		}
159 	}
160 
161 	sprintf(node, "node%d", hive->number_devices);
162 	/* Create sysfs link form the hive folder to yourself */
163 	ret = sysfs_create_link(hive->kobj, &adev->dev->kobj, node);
164 	if (ret) {
165 		dev_err(adev->dev, "XGMI: Failed to create link from hive info");
166 		goto remove_link;
167 	}
168 
169 	goto success;
170 
171 
172 remove_link:
173 	sysfs_remove_link(&adev->dev->kobj, adev->ddev->unique);
174 
175 remove_file:
176 	device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
177 
178 success:
179 	return ret;
180 }
181 
182 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev,
183 					  struct amdgpu_hive_info *hive)
184 {
185 	device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
186 	sysfs_remove_link(&adev->dev->kobj, adev->ddev->unique);
187 	sysfs_remove_link(hive->kobj, adev->ddev->unique);
188 }
189 
190 
191 
192 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev, int lock)
193 {
194 	int i;
195 	struct amdgpu_hive_info *tmp;
196 
197 	if (!adev->gmc.xgmi.hive_id)
198 		return NULL;
199 
200 	mutex_lock(&xgmi_mutex);
201 
202 	for (i = 0 ; i < hive_count; ++i) {
203 		tmp = &xgmi_hives[i];
204 		if (tmp->hive_id == adev->gmc.xgmi.hive_id) {
205 			if (lock)
206 				mutex_lock(&tmp->hive_lock);
207 			mutex_unlock(&xgmi_mutex);
208 			return tmp;
209 		}
210 	}
211 	if (i >= AMDGPU_MAX_XGMI_HIVE) {
212 		mutex_unlock(&xgmi_mutex);
213 		return NULL;
214 	}
215 
216 	/* initialize new hive if not exist */
217 	tmp = &xgmi_hives[hive_count++];
218 
219 	if (amdgpu_xgmi_sysfs_create(adev, tmp)) {
220 		mutex_unlock(&xgmi_mutex);
221 		return NULL;
222 	}
223 
224 	tmp->adev = adev;
225 	tmp->hive_id = adev->gmc.xgmi.hive_id;
226 	INIT_LIST_HEAD(&tmp->device_list);
227 	mutex_init(&tmp->hive_lock);
228 	mutex_init(&tmp->reset_lock);
229 
230 	if (lock)
231 		mutex_lock(&tmp->hive_lock);
232 	tmp->pstate = -1;
233 	mutex_unlock(&xgmi_mutex);
234 
235 	return tmp;
236 }
237 
238 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate)
239 {
240 	int ret = 0;
241 	struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev, 0);
242 
243 	if (!hive)
244 		return 0;
245 
246 	if (hive->pstate == pstate)
247 		return 0;
248 
249 	dev_dbg(adev->dev, "Set xgmi pstate %d.\n", pstate);
250 
251 	if (is_support_sw_smu(adev))
252 		ret = smu_set_xgmi_pstate(&adev->smu, pstate);
253 	if (ret)
254 		dev_err(adev->dev,
255 			"XGMI: Set pstate failure on device %llx, hive %llx, ret %d",
256 			adev->gmc.xgmi.node_id,
257 			adev->gmc.xgmi.hive_id, ret);
258 
259 	return ret;
260 }
261 
262 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev)
263 {
264 	int ret = -EINVAL;
265 
266 	/* Each psp need to set the latest topology */
267 	ret = psp_xgmi_set_topology_info(&adev->psp,
268 					 hive->number_devices,
269 					 &adev->psp.xgmi_context.top_info);
270 	if (ret)
271 		dev_err(adev->dev,
272 			"XGMI: Set topology failure on device %llx, hive %llx, ret %d",
273 			adev->gmc.xgmi.node_id,
274 			adev->gmc.xgmi.hive_id, ret);
275 
276 	return ret;
277 }
278 
279 
280 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev,
281 		struct amdgpu_device *peer_adev)
282 {
283 	struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
284 	int i;
285 
286 	for (i = 0 ; i < top->num_nodes; ++i)
287 		if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
288 			return top->nodes[i].num_hops;
289 	return	-EINVAL;
290 }
291 
292 int amdgpu_xgmi_add_device(struct amdgpu_device *adev)
293 {
294 	struct psp_xgmi_topology_info *top_info;
295 	struct amdgpu_hive_info *hive;
296 	struct amdgpu_xgmi	*entry;
297 	struct amdgpu_device *tmp_adev = NULL;
298 
299 	int count = 0, ret = -EINVAL;
300 
301 	if (!adev->gmc.xgmi.supported)
302 		return 0;
303 
304 	ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id);
305 	if (ret) {
306 		dev_err(adev->dev,
307 			"XGMI: Failed to get node id\n");
308 		return ret;
309 	}
310 
311 	ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id);
312 	if (ret) {
313 		dev_err(adev->dev,
314 			"XGMI: Failed to get hive id\n");
315 		return ret;
316 	}
317 
318 	hive = amdgpu_get_xgmi_hive(adev, 1);
319 	if (!hive) {
320 		ret = -EINVAL;
321 		dev_err(adev->dev,
322 			"XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n",
323 			adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id);
324 		goto exit;
325 	}
326 
327 	top_info = &adev->psp.xgmi_context.top_info;
328 
329 	list_add_tail(&adev->gmc.xgmi.head, &hive->device_list);
330 	list_for_each_entry(entry, &hive->device_list, head)
331 		top_info->nodes[count++].node_id = entry->node_id;
332 	top_info->num_nodes = count;
333 	hive->number_devices = count;
334 
335 	list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
336 		/* update node list for other device in the hive */
337 		if (tmp_adev != adev) {
338 			top_info = &tmp_adev->psp.xgmi_context.top_info;
339 			top_info->nodes[count - 1].node_id = adev->gmc.xgmi.node_id;
340 			top_info->num_nodes = count;
341 		}
342 		ret = amdgpu_xgmi_update_topology(hive, tmp_adev);
343 		if (ret)
344 			goto exit;
345 	}
346 
347 	/* get latest topology info for each device from psp */
348 	list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
349 		ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
350 				&tmp_adev->psp.xgmi_context.top_info);
351 		if (ret) {
352 			dev_err(tmp_adev->dev,
353 				"XGMI: Get topology failure on device %llx, hive %llx, ret %d",
354 				tmp_adev->gmc.xgmi.node_id,
355 				tmp_adev->gmc.xgmi.hive_id, ret);
356 			/* To do : continue with some node failed or disable the whole hive */
357 			goto exit;
358 		}
359 	}
360 
361 	if (!ret)
362 		ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive);
363 
364 
365 	mutex_unlock(&hive->hive_lock);
366 exit:
367 	if (!ret)
368 		dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n",
369 			 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id);
370 	else
371 		dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n",
372 			adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id,
373 			ret);
374 
375 	return ret;
376 }
377 
378 void amdgpu_xgmi_remove_device(struct amdgpu_device *adev)
379 {
380 	struct amdgpu_hive_info *hive;
381 
382 	if (!adev->gmc.xgmi.supported)
383 		return;
384 
385 	hive = amdgpu_get_xgmi_hive(adev, 1);
386 	if (!hive)
387 		return;
388 
389 	if (!(hive->number_devices--)) {
390 		amdgpu_xgmi_sysfs_destroy(adev, hive);
391 		mutex_destroy(&hive->hive_lock);
392 		mutex_destroy(&hive->reset_lock);
393 	} else {
394 		amdgpu_xgmi_sysfs_rem_dev_info(adev, hive);
395 		mutex_unlock(&hive->hive_lock);
396 	}
397 }
398