xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdgpu/amdgpu_xgmi.c (revision e4a580baf2598beeaae98d953ac7635b8700b80c)
1 /*	$NetBSD: amdgpu_xgmi.c,v 1.3 2021/12/19 10:59:01 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2018 Advanced Micro Devices, Inc.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  *
25  */
26 #include <sys/cdefs.h>
27 __KERNEL_RCSID(0, "$NetBSD: amdgpu_xgmi.c,v 1.3 2021/12/19 10:59:01 riastradh Exp $");
28 
29 #include <linux/list.h>
30 #include "amdgpu.h"
31 #include "amdgpu_xgmi.h"
32 #include "amdgpu_smu.h"
33 #include "amdgpu_ras.h"
34 #include "df/df_3_6_offset.h"
35 
36 static DEFINE_MUTEX(xgmi_mutex);
37 
38 #define AMDGPU_MAX_XGMI_HIVE			8
39 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE		4
40 
41 static struct amdgpu_hive_info xgmi_hives[AMDGPU_MAX_XGMI_HIVE];
42 static unsigned hive_count = 0;
43 
amdgpu_xgmi_hive_try_lock(struct amdgpu_hive_info * hive)44 void *amdgpu_xgmi_hive_try_lock(struct amdgpu_hive_info *hive)
45 {
46 	return &hive->device_list;
47 }
48 
49 /**
50  * DOC: AMDGPU XGMI Support
51  *
52  * XGMI is a high speed interconnect that joins multiple GPU cards
53  * into a homogeneous memory space that is organized by a collective
54  * hive ID and individual node IDs, both of which are 64-bit numbers.
55  *
56  * The file xgmi_device_id contains the unique per GPU device ID and
57  * is stored in the /sys/class/drm/card${cardno}/device/ directory.
58  *
59  * Inside the device directory a sub-directory 'xgmi_hive_info' is
60  * created which contains the hive ID and the list of nodes.
61  *
62  * The hive ID is stored in:
63  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id
64  *
65  * The node information is stored in numbered directories:
66  *   /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id
67  *
68  * Each device has their own xgmi_hive_info direction with a mirror
69  * set of node sub-directories.
70  *
71  * The XGMI memory space is built by contiguously adding the power of
72  * two padded VRAM space from each node to each other.
73  *
74  */
75 
76 
amdgpu_xgmi_show_hive_id(struct device * dev,struct device_attribute * attr,char * buf)77 static ssize_t amdgpu_xgmi_show_hive_id(struct device *dev,
78 		struct device_attribute *attr, char *buf)
79 {
80 	struct amdgpu_hive_info *hive =
81 			container_of(attr, struct amdgpu_hive_info, dev_attr);
82 
83 	return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id);
84 }
85 
amdgpu_xgmi_sysfs_create(struct amdgpu_device * adev,struct amdgpu_hive_info * hive)86 static int amdgpu_xgmi_sysfs_create(struct amdgpu_device *adev,
87 				    struct amdgpu_hive_info *hive)
88 {
89 #ifdef CONFIG_SYSFS
90 	int ret = 0;
91 
92 	if (WARN_ON(hive->kobj))
93 		return -EINVAL;
94 
95 	hive->kobj = kobject_create_and_add("xgmi_hive_info", &adev->dev->kobj);
96 	if (!hive->kobj) {
97 		dev_err(adev->dev, "XGMI: Failed to allocate sysfs entry!\n");
98 		return -EINVAL;
99 	}
100 
101 	hive->dev_attr = (struct device_attribute) {
102 		.attr = {
103 			.name = "xgmi_hive_id",
104 			.mode = S_IRUGO,
105 
106 		},
107 		.show = amdgpu_xgmi_show_hive_id,
108 	};
109 
110 	ret = sysfs_create_file(hive->kobj, &hive->dev_attr.attr);
111 	if (ret) {
112 		dev_err(adev->dev, "XGMI: Failed to create device file xgmi_hive_id\n");
113 		kobject_del(hive->kobj);
114 		kobject_put(hive->kobj);
115 		hive->kobj = NULL;
116 	}
117 
118 	return ret;
119 #endif
120 }
121 
amdgpu_xgmi_sysfs_destroy(struct amdgpu_device * adev,struct amdgpu_hive_info * hive)122 static void amdgpu_xgmi_sysfs_destroy(struct amdgpu_device *adev,
123 				    struct amdgpu_hive_info *hive)
124 {
125 #ifdef CONFIG_SYSFS
126 	sysfs_remove_file(hive->kobj, &hive->dev_attr.attr);
127 	kobject_del(hive->kobj);
128 	kobject_put(hive->kobj);
129 	hive->kobj = NULL;
130 #endif
131 }
132 
amdgpu_xgmi_show_device_id(struct device * dev,struct device_attribute * attr,char * buf)133 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev,
134 				     struct device_attribute *attr,
135 				     char *buf)
136 {
137 	struct drm_device *ddev = dev_get_drvdata(dev);
138 	struct amdgpu_device *adev = ddev->dev_private;
139 
140 	return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.xgmi.node_id);
141 
142 }
143 
144 #define AMDGPU_XGMI_SET_FICAA(o)	((o) | 0x456801)
amdgpu_xgmi_show_error(struct device * dev,struct device_attribute * attr,char * buf)145 static ssize_t amdgpu_xgmi_show_error(struct device *dev,
146 				      struct device_attribute *attr,
147 				      char *buf)
148 {
149 	struct drm_device *ddev = dev_get_drvdata(dev);
150 	struct amdgpu_device *adev = ddev->dev_private;
151 	uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in;
152 	uint64_t fica_out;
153 	unsigned int error_count = 0;
154 
155 	ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200);
156 	ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208);
157 
158 	fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in);
159 	if (fica_out != 0x1f)
160 		pr_err("xGMI error counters not enabled!\n");
161 
162 	fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in);
163 
164 	if ((fica_out & 0xffff) == 2)
165 		error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63);
166 
167 	adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0);
168 
169 	return snprintf(buf, PAGE_SIZE, "%d\n", error_count);
170 }
171 
172 
173 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL);
174 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL);
175 
amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device * adev,struct amdgpu_hive_info * hive)176 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev,
177 					 struct amdgpu_hive_info *hive)
178 {
179 #ifdef CONFIG_SYSFS
180 	int ret = 0;
181 	char node[10] = { 0 };
182 
183 	/* Create xgmi device id file */
184 	ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id);
185 	if (ret) {
186 		dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n");
187 		return ret;
188 	}
189 
190 	/* Create xgmi error file */
191 	ret = device_create_file(adev->dev, &dev_attr_xgmi_error);
192 	if (ret)
193 		pr_err("failed to create xgmi_error\n");
194 
195 
196 	/* Create sysfs link to hive info folder on the first device */
197 	if (adev != hive->adev) {
198 		ret = sysfs_create_link(&adev->dev->kobj, hive->kobj,
199 					"xgmi_hive_info");
200 		if (ret) {
201 			dev_err(adev->dev, "XGMI: Failed to create link to hive info");
202 			goto remove_file;
203 		}
204 	}
205 
206 	sprintf(node, "node%d", hive->number_devices);
207 	/* Create sysfs link form the hive folder to yourself */
208 	ret = sysfs_create_link(hive->kobj, &adev->dev->kobj, node);
209 	if (ret) {
210 		dev_err(adev->dev, "XGMI: Failed to create link from hive info");
211 		goto remove_link;
212 	}
213 
214 	goto success;
215 
216 
217 remove_link:
218 	sysfs_remove_link(&adev->dev->kobj, adev->ddev->unique);
219 
220 remove_file:
221 	device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
222 
223 success:
224 	return ret;
225 #endif
226 }
227 
amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device * adev,struct amdgpu_hive_info * hive)228 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev,
229 					  struct amdgpu_hive_info *hive)
230 {
231 #ifdef CONFIG_SYSFS
232 	device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
233 	sysfs_remove_link(&adev->dev->kobj, adev->ddev->unique);
234 	sysfs_remove_link(hive->kobj, adev->ddev->unique);
235 #endif
236 }
237 
238 
239 
amdgpu_get_xgmi_hive(struct amdgpu_device * adev,int lock)240 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev, int lock)
241 {
242 	int i;
243 	struct amdgpu_hive_info *tmp;
244 
245 	if (!adev->gmc.xgmi.hive_id)
246 		return NULL;
247 
248 	mutex_lock(&xgmi_mutex);
249 
250 	for (i = 0 ; i < hive_count; ++i) {
251 		tmp = &xgmi_hives[i];
252 		if (tmp->hive_id == adev->gmc.xgmi.hive_id) {
253 			if (lock)
254 				mutex_lock(&tmp->hive_lock);
255 			mutex_unlock(&xgmi_mutex);
256 			return tmp;
257 		}
258 	}
259 	if (i >= AMDGPU_MAX_XGMI_HIVE) {
260 		mutex_unlock(&xgmi_mutex);
261 		return NULL;
262 	}
263 
264 	/* initialize new hive if not exist */
265 	tmp = &xgmi_hives[hive_count++];
266 
267 	if (amdgpu_xgmi_sysfs_create(adev, tmp)) {
268 		mutex_unlock(&xgmi_mutex);
269 		return NULL;
270 	}
271 
272 	tmp->adev = adev;
273 	tmp->hive_id = adev->gmc.xgmi.hive_id;
274 	INIT_LIST_HEAD(&tmp->device_list);
275 	mutex_init(&tmp->hive_lock);
276 	mutex_init(&tmp->reset_lock);
277 	task_barrier_init(&tmp->tb);
278 
279 	if (lock)
280 		mutex_lock(&tmp->hive_lock);
281 	tmp->pstate = -1;
282 	mutex_unlock(&xgmi_mutex);
283 
284 	return tmp;
285 }
286 
amdgpu_xgmi_set_pstate(struct amdgpu_device * adev,int pstate)287 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate)
288 {
289 	int ret = 0;
290 	struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev, 0);
291 	struct amdgpu_device *tmp_adev;
292 	bool update_hive_pstate = true;
293 	bool is_high_pstate = pstate && adev->asic_type == CHIP_VEGA20;
294 
295 	if (!hive)
296 		return 0;
297 
298 	mutex_lock(&hive->hive_lock);
299 
300 	if (hive->pstate == pstate) {
301 		adev->pstate = is_high_pstate ? pstate : adev->pstate;
302 		goto out;
303 	}
304 
305 	dev_dbg(adev->dev, "Set xgmi pstate %d.\n", pstate);
306 
307 	ret = amdgpu_dpm_set_xgmi_pstate(adev, pstate);
308 	if (ret) {
309 		dev_err(adev->dev,
310 			"XGMI: Set pstate failure on device %llx, hive %llx, ret %d",
311 			adev->gmc.xgmi.node_id,
312 			adev->gmc.xgmi.hive_id, ret);
313 		goto out;
314 	}
315 
316 	/* Update device pstate */
317 	adev->pstate = pstate;
318 
319 	/*
320 	 * Update the hive pstate only all devices of the hive
321 	 * are in the same pstate
322 	 */
323 	list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
324 		if (tmp_adev->pstate != adev->pstate) {
325 			update_hive_pstate = false;
326 			break;
327 		}
328 	}
329 	if (update_hive_pstate || is_high_pstate)
330 		hive->pstate = pstate;
331 
332 out:
333 	mutex_unlock(&hive->hive_lock);
334 
335 	return ret;
336 }
337 
amdgpu_xgmi_update_topology(struct amdgpu_hive_info * hive,struct amdgpu_device * adev)338 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev)
339 {
340 	int ret = -EINVAL;
341 
342 	/* Each psp need to set the latest topology */
343 	ret = psp_xgmi_set_topology_info(&adev->psp,
344 					 hive->number_devices,
345 					 &adev->psp.xgmi_context.top_info);
346 	if (ret)
347 		dev_err(adev->dev,
348 			"XGMI: Set topology failure on device %llx, hive %llx, ret %d",
349 			adev->gmc.xgmi.node_id,
350 			adev->gmc.xgmi.hive_id, ret);
351 
352 	return ret;
353 }
354 
355 
amdgpu_xgmi_get_hops_count(struct amdgpu_device * adev,struct amdgpu_device * peer_adev)356 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev,
357 		struct amdgpu_device *peer_adev)
358 {
359 	struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
360 	int i;
361 
362 	for (i = 0 ; i < top->num_nodes; ++i)
363 		if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
364 			return top->nodes[i].num_hops;
365 	return	-EINVAL;
366 }
367 
amdgpu_xgmi_add_device(struct amdgpu_device * adev)368 int amdgpu_xgmi_add_device(struct amdgpu_device *adev)
369 {
370 	struct psp_xgmi_topology_info *top_info;
371 	struct amdgpu_hive_info *hive;
372 	struct amdgpu_xgmi	*entry;
373 	struct amdgpu_device *tmp_adev = NULL;
374 
375 	int count = 0, ret = 0;
376 
377 	if (!adev->gmc.xgmi.supported)
378 		return 0;
379 
380 	if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
381 		ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id);
382 		if (ret) {
383 			dev_err(adev->dev,
384 				"XGMI: Failed to get hive id\n");
385 			return ret;
386 		}
387 
388 		ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id);
389 		if (ret) {
390 			dev_err(adev->dev,
391 				"XGMI: Failed to get node id\n");
392 			return ret;
393 		}
394 	} else {
395 		adev->gmc.xgmi.hive_id = 16;
396 		adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16;
397 	}
398 
399 	hive = amdgpu_get_xgmi_hive(adev, 1);
400 	if (!hive) {
401 		ret = -EINVAL;
402 		dev_err(adev->dev,
403 			"XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n",
404 			adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id);
405 		goto exit;
406 	}
407 
408 	/* Set default device pstate */
409 	adev->pstate = -1;
410 
411 	top_info = &adev->psp.xgmi_context.top_info;
412 
413 	list_add_tail(&adev->gmc.xgmi.head, &hive->device_list);
414 	list_for_each_entry(entry, &hive->device_list, head)
415 		top_info->nodes[count++].node_id = entry->node_id;
416 	top_info->num_nodes = count;
417 	hive->number_devices = count;
418 
419 	task_barrier_add_task(&hive->tb);
420 
421 	if (amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
422 		list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
423 			/* update node list for other device in the hive */
424 			if (tmp_adev != adev) {
425 				top_info = &tmp_adev->psp.xgmi_context.top_info;
426 				top_info->nodes[count - 1].node_id =
427 					adev->gmc.xgmi.node_id;
428 				top_info->num_nodes = count;
429 			}
430 			ret = amdgpu_xgmi_update_topology(hive, tmp_adev);
431 			if (ret)
432 				goto exit;
433 		}
434 
435 		/* get latest topology info for each device from psp */
436 		list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
437 			ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
438 					&tmp_adev->psp.xgmi_context.top_info);
439 			if (ret) {
440 				dev_err(tmp_adev->dev,
441 					"XGMI: Get topology failure on device %llx, hive %llx, ret %d",
442 					tmp_adev->gmc.xgmi.node_id,
443 					tmp_adev->gmc.xgmi.hive_id, ret);
444 				/* To do : continue with some node failed or disable the whole hive */
445 				goto exit;
446 			}
447 		}
448 	}
449 
450 	if (!ret)
451 		ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive);
452 
453 
454 	mutex_unlock(&hive->hive_lock);
455 exit:
456 	if (!ret)
457 		dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n",
458 			 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id);
459 	else
460 		dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n",
461 			adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id,
462 			ret);
463 
464 	return ret;
465 }
466 
amdgpu_xgmi_remove_device(struct amdgpu_device * adev)467 void amdgpu_xgmi_remove_device(struct amdgpu_device *adev)
468 {
469 	struct amdgpu_hive_info *hive;
470 
471 	if (!adev->gmc.xgmi.supported)
472 		return;
473 
474 	hive = amdgpu_get_xgmi_hive(adev, 1);
475 	if (!hive)
476 		return;
477 
478 	if (!(hive->number_devices--)) {
479 		amdgpu_xgmi_sysfs_destroy(adev, hive);
480 		task_barrier_destroy(&tmp->tb);
481 		mutex_destroy(&hive->hive_lock);
482 		mutex_destroy(&hive->reset_lock);
483 	} else {
484 		task_barrier_rem_task(&hive->tb);
485 		amdgpu_xgmi_sysfs_rem_dev_info(adev, hive);
486 		mutex_unlock(&hive->hive_lock);
487 	}
488 }
489 
amdgpu_xgmi_ras_late_init(struct amdgpu_device * adev)490 int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev)
491 {
492 	int r;
493 	struct ras_ih_if ih_info = {
494 		.cb = NULL,
495 	};
496 	struct ras_fs_if fs_info = {
497 		.sysfs_name = "xgmi_wafl_err_count",
498 		.debugfs_name = "xgmi_wafl_err_inject",
499 	};
500 
501 	if (!adev->gmc.xgmi.supported ||
502 	    adev->gmc.xgmi.num_physical_nodes == 0)
503 		return 0;
504 
505 	if (!adev->gmc.xgmi.ras_if) {
506 		adev->gmc.xgmi.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL);
507 		if (!adev->gmc.xgmi.ras_if)
508 			return -ENOMEM;
509 		adev->gmc.xgmi.ras_if->block = AMDGPU_RAS_BLOCK__XGMI_WAFL;
510 		adev->gmc.xgmi.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
511 		adev->gmc.xgmi.ras_if->sub_block_index = 0;
512 		strcpy(adev->gmc.xgmi.ras_if->name, "xgmi_wafl");
513 	}
514 	ih_info.head = fs_info.head = *adev->gmc.xgmi.ras_if;
515 	r = amdgpu_ras_late_init(adev, adev->gmc.xgmi.ras_if,
516 				 &fs_info, &ih_info);
517 	if (r || !amdgpu_ras_is_supported(adev, adev->gmc.xgmi.ras_if->block)) {
518 		kfree(adev->gmc.xgmi.ras_if);
519 		adev->gmc.xgmi.ras_if = NULL;
520 	}
521 
522 	return r;
523 }
524 
amdgpu_xgmi_ras_fini(struct amdgpu_device * adev)525 void amdgpu_xgmi_ras_fini(struct amdgpu_device *adev)
526 {
527 	if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL) &&
528 			adev->gmc.xgmi.ras_if) {
529 		struct ras_common_if *ras_if = adev->gmc.xgmi.ras_if;
530 		struct ras_ih_if ih_info = {
531 			.cb = NULL,
532 		};
533 
534 		amdgpu_ras_late_fini(adev, ras_if, &ih_info);
535 		kfree(ras_if);
536 	}
537 }
538