2 * Copyright 2018 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
24 #include <linux/list.h>
26 #include "amdgpu_xgmi.h"
27 #include "amdgpu_smu.h"
28 #include "amdgpu_ras.h"
30 #include "df/df_3_6_offset.h"
31 #include "xgmi/xgmi_4_0_0_smn.h"
32 #include "xgmi/xgmi_4_0_0_sh_mask.h"
33 #include "wafl/wafl2_4_0_0_smn.h"
34 #include "wafl/wafl2_4_0_0_sh_mask.h"
36 static DEFINE_MUTEX(xgmi_mutex);
38 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE 4
40 static LIST_HEAD(xgmi_hive_list);
42 static const int xgmi_pcs_err_status_reg_vg20[] = {
43 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS,
44 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000,
47 static const int wafl_pcs_err_status_reg_vg20[] = {
48 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS,
49 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000,
52 static const int xgmi_pcs_err_status_reg_arct[] = {
53 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS,
54 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000,
55 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000,
56 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000,
57 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000,
58 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000,
62 static const int wafl_pcs_err_status_reg_arct[] = {
63 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS,
64 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000,
67 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = {
68 {"XGMI PCS DataLossErr",
69 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)},
70 {"XGMI PCS TrainingErr",
71 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)},
73 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)},
74 {"XGMI PCS BERExceededErr",
75 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)},
76 {"XGMI PCS TxMetaDataErr",
77 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)},
78 {"XGMI PCS ReplayBufParityErr",
79 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)},
80 {"XGMI PCS DataParityErr",
81 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)},
82 {"XGMI PCS ReplayFifoOverflowErr",
83 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
84 {"XGMI PCS ReplayFifoUnderflowErr",
85 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
86 {"XGMI PCS ElasticFifoOverflowErr",
87 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
88 {"XGMI PCS DeskewErr",
89 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)},
90 {"XGMI PCS DataStartupLimitErr",
91 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)},
92 {"XGMI PCS FCInitTimeoutErr",
93 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)},
94 {"XGMI PCS RecoveryTimeoutErr",
95 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
96 {"XGMI PCS ReadySerialTimeoutErr",
97 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
98 {"XGMI PCS ReadySerialAttemptErr",
99 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
100 {"XGMI PCS RecoveryAttemptErr",
101 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)},
102 {"XGMI PCS RecoveryRelockAttemptErr",
103 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
106 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = {
107 {"WAFL PCS DataLossErr",
108 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)},
109 {"WAFL PCS TrainingErr",
110 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)},
112 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)},
113 {"WAFL PCS BERExceededErr",
114 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)},
115 {"WAFL PCS TxMetaDataErr",
116 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)},
117 {"WAFL PCS ReplayBufParityErr",
118 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)},
119 {"WAFL PCS DataParityErr",
120 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)},
121 {"WAFL PCS ReplayFifoOverflowErr",
122 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
123 {"WAFL PCS ReplayFifoUnderflowErr",
124 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
125 {"WAFL PCS ElasticFifoOverflowErr",
126 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
127 {"WAFL PCS DeskewErr",
128 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)},
129 {"WAFL PCS DataStartupLimitErr",
130 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)},
131 {"WAFL PCS FCInitTimeoutErr",
132 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)},
133 {"WAFL PCS RecoveryTimeoutErr",
134 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
135 {"WAFL PCS ReadySerialTimeoutErr",
136 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
137 {"WAFL PCS ReadySerialAttemptErr",
138 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
139 {"WAFL PCS RecoveryAttemptErr",
140 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)},
141 {"WAFL PCS RecoveryRelockAttemptErr",
142 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
146 * DOC: AMDGPU XGMI Support
148 * XGMI is a high speed interconnect that joins multiple GPU cards
149 * into a homogeneous memory space that is organized by a collective
150 * hive ID and individual node IDs, both of which are 64-bit numbers.
152 * The file xgmi_device_id contains the unique per GPU device ID and
153 * is stored in the /sys/class/drm/card${cardno}/device/ directory.
155 * Inside the device directory a sub-directory 'xgmi_hive_info' is
156 * created which contains the hive ID and the list of nodes.
158 * The hive ID is stored in:
159 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id
161 * The node information is stored in numbered directories:
162 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id
164 * Each device has their own xgmi_hive_info direction with a mirror
165 * set of node sub-directories.
167 * The XGMI memory space is built by contiguously adding the power of
168 * two padded VRAM space from each node to each other.
172 static struct attribute amdgpu_xgmi_hive_id = {
173 .name = "xgmi_hive_id",
177 static struct attribute *amdgpu_xgmi_hive_attrs[] = {
178 &amdgpu_xgmi_hive_id,
182 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj,
183 struct attribute *attr, char *buf)
185 struct amdgpu_hive_info *hive = container_of(
186 kobj, struct amdgpu_hive_info, kobj);
188 if (attr == &amdgpu_xgmi_hive_id)
189 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id);
194 static void amdgpu_xgmi_hive_release(struct kobject *kobj)
196 struct amdgpu_hive_info *hive = container_of(
197 kobj, struct amdgpu_hive_info, kobj);
199 mutex_destroy(&hive->hive_lock);
203 static const struct sysfs_ops amdgpu_xgmi_hive_ops = {
204 .show = amdgpu_xgmi_show_attrs,
207 struct kobj_type amdgpu_xgmi_hive_type = {
208 .release = amdgpu_xgmi_hive_release,
209 .sysfs_ops = &amdgpu_xgmi_hive_ops,
210 .default_attrs = amdgpu_xgmi_hive_attrs,
213 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev,
214 struct device_attribute *attr,
217 struct drm_device *ddev = dev_get_drvdata(dev);
218 struct amdgpu_device *adev = drm_to_adev(ddev);
220 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.xgmi.node_id);
224 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801)
225 static ssize_t amdgpu_xgmi_show_error(struct device *dev,
226 struct device_attribute *attr,
229 struct drm_device *ddev = dev_get_drvdata(dev);
230 struct amdgpu_device *adev = drm_to_adev(ddev);
231 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in;
233 unsigned int error_count = 0;
235 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200);
236 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208);
238 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in);
239 if (fica_out != 0x1f)
240 pr_err("xGMI error counters not enabled!\n");
242 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in);
244 if ((fica_out & 0xffff) == 2)
245 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63);
247 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0);
249 return snprintf(buf, PAGE_SIZE, "%u\n", error_count);
253 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL);
254 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL);
256 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev,
257 struct amdgpu_hive_info *hive)
260 char node[10] = { 0 };
262 /* Create xgmi device id file */
263 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id);
265 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n");
269 /* Create xgmi error file */
270 ret = device_create_file(adev->dev, &dev_attr_xgmi_error);
272 pr_err("failed to create xgmi_error\n");
275 /* Create sysfs link to hive info folder on the first device */
276 if (hive->kobj.parent != (&adev->dev->kobj)) {
277 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj,
280 dev_err(adev->dev, "XGMI: Failed to create link to hive info");
285 sprintf(node, "node%d", atomic_read(&hive->number_devices));
286 /* Create sysfs link form the hive folder to yourself */
287 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node);
289 dev_err(adev->dev, "XGMI: Failed to create link from hive info");
297 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique);
300 device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
306 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev,
307 struct amdgpu_hive_info *hive)
310 memset(node, 0, sizeof(node));
312 device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
313 device_remove_file(adev->dev, &dev_attr_xgmi_error);
315 if (hive->kobj.parent != (&adev->dev->kobj))
316 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info");
318 sprintf(node, "node%d", atomic_read(&hive->number_devices));
319 sysfs_remove_link(&hive->kobj, node);
325 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev)
327 struct amdgpu_hive_info *hive = NULL;
330 if (!adev->gmc.xgmi.hive_id)
334 kobject_get(&adev->hive->kobj);
338 mutex_lock(&xgmi_mutex);
340 list_for_each_entry(hive, &xgmi_hive_list, node) {
341 if (hive->hive_id == adev->gmc.xgmi.hive_id)
345 hive = kzalloc(sizeof(*hive), GFP_KERNEL);
347 dev_err(adev->dev, "XGMI: allocation failed\n");
352 /* initialize new hive if not exist */
353 ret = kobject_init_and_add(&hive->kobj,
354 &amdgpu_xgmi_hive_type,
356 "%s", "xgmi_hive_info");
358 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n");
364 hive->hive_id = adev->gmc.xgmi.hive_id;
365 INIT_LIST_HEAD(&hive->device_list);
366 INIT_LIST_HEAD(&hive->node);
367 mutex_init(&hive->hive_lock);
368 atomic_set(&hive->in_reset, 0);
369 atomic_set(&hive->number_devices, 0);
370 task_barrier_init(&hive->tb);
371 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN;
372 hive->hi_req_gpu = NULL;
374 * hive pstate on boot is high in vega20 so we have to go to low
375 * pstate on after boot.
377 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE;
378 list_add_tail(&hive->node, &xgmi_hive_list);
382 kobject_get(&hive->kobj);
383 mutex_unlock(&xgmi_mutex);
387 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive)
390 kobject_put(&hive->kobj);
393 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate)
396 struct amdgpu_hive_info *hive;
397 struct amdgpu_device *request_adev;
398 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20;
401 hive = amdgpu_get_xgmi_hive(adev);
405 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev;
406 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN;
407 amdgpu_put_xgmi_hive(hive);
408 /* fw bug so temporarily disable pstate switching */
411 if (!hive || adev->asic_type != CHIP_VEGA20)
414 mutex_lock(&hive->hive_lock);
417 hive->hi_req_count++;
419 hive->hi_req_count--;
422 * Vega20 only needs single peer to request pstate high for the hive to
423 * go high but all peers must request pstate low for the hive to go low
425 if (hive->pstate == pstate ||
426 (!is_hi_req && hive->hi_req_count && !init_low))
429 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate);
431 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate);
433 dev_err(request_adev->dev,
434 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d",
435 request_adev->gmc.xgmi.node_id,
436 request_adev->gmc.xgmi.hive_id, ret);
441 hive->pstate = hive->hi_req_count ?
442 hive->pstate : AMDGPU_XGMI_PSTATE_MIN;
444 hive->pstate = pstate;
445 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ?
449 mutex_unlock(&hive->hive_lock);
453 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev)
457 /* Each psp need to set the latest topology */
458 ret = psp_xgmi_set_topology_info(&adev->psp,
459 atomic_read(&hive->number_devices),
460 &adev->psp.xgmi_context.top_info);
463 "XGMI: Set topology failure on device %llx, hive %llx, ret %d",
464 adev->gmc.xgmi.node_id,
465 adev->gmc.xgmi.hive_id, ret);
472 * NOTE psp_xgmi_node_info.num_hops layout is as follows:
473 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved)
474 * num_hops[5:3] = reserved
475 * num_hops[2:0] = number of hops
477 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev,
478 struct amdgpu_device *peer_adev)
480 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
481 uint8_t num_hops_mask = 0x7;
484 for (i = 0 ; i < top->num_nodes; ++i)
485 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
486 return top->nodes[i].num_hops & num_hops_mask;
490 int amdgpu_xgmi_add_device(struct amdgpu_device *adev)
492 struct psp_xgmi_topology_info *top_info;
493 struct amdgpu_hive_info *hive;
494 struct amdgpu_xgmi *entry;
495 struct amdgpu_device *tmp_adev = NULL;
497 int count = 0, ret = 0;
499 if (!adev->gmc.xgmi.supported)
502 if (!adev->gmc.xgmi.pending_reset &&
503 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
504 ret = psp_xgmi_initialize(&adev->psp);
507 "XGMI: Failed to initialize xgmi session\n");
511 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id);
514 "XGMI: Failed to get hive id\n");
518 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id);
521 "XGMI: Failed to get node id\n");
525 adev->gmc.xgmi.hive_id = 16;
526 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16;
529 hive = amdgpu_get_xgmi_hive(adev);
533 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n",
534 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id);
537 mutex_lock(&hive->hive_lock);
539 top_info = &adev->psp.xgmi_context.top_info;
541 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list);
542 list_for_each_entry(entry, &hive->device_list, head)
543 top_info->nodes[count++].node_id = entry->node_id;
544 top_info->num_nodes = count;
545 atomic_set(&hive->number_devices, count);
547 task_barrier_add_task(&hive->tb);
549 if (!adev->gmc.xgmi.pending_reset &&
550 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
551 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
552 /* update node list for other device in the hive */
553 if (tmp_adev != adev) {
554 top_info = &tmp_adev->psp.xgmi_context.top_info;
555 top_info->nodes[count - 1].node_id =
556 adev->gmc.xgmi.node_id;
557 top_info->num_nodes = count;
559 ret = amdgpu_xgmi_update_topology(hive, tmp_adev);
564 /* get latest topology info for each device from psp */
565 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
566 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
567 &tmp_adev->psp.xgmi_context.top_info);
569 dev_err(tmp_adev->dev,
570 "XGMI: Get topology failure on device %llx, hive %llx, ret %d",
571 tmp_adev->gmc.xgmi.node_id,
572 tmp_adev->gmc.xgmi.hive_id, ret);
573 /* To do : continue with some node failed or disable the whole hive */
579 if (!ret && !adev->gmc.xgmi.pending_reset)
580 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive);
583 mutex_unlock(&hive->hive_lock);
587 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n",
588 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id);
590 amdgpu_put_xgmi_hive(hive);
591 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n",
592 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id,
599 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev)
601 struct amdgpu_hive_info *hive = adev->hive;
603 if (!adev->gmc.xgmi.supported)
609 mutex_lock(&hive->hive_lock);
610 task_barrier_rem_task(&hive->tb);
611 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive);
612 if (hive->hi_req_gpu == adev)
613 hive->hi_req_gpu = NULL;
614 list_del(&adev->gmc.xgmi.head);
615 mutex_unlock(&hive->hive_lock);
617 amdgpu_put_xgmi_hive(hive);
620 if (atomic_dec_return(&hive->number_devices) == 0) {
621 /* Remove the hive from global hive list */
622 mutex_lock(&xgmi_mutex);
623 list_del(&hive->node);
624 mutex_unlock(&xgmi_mutex);
626 amdgpu_put_xgmi_hive(hive);
629 return psp_xgmi_terminate(&adev->psp);
632 int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev)
635 struct ras_ih_if ih_info = {
638 struct ras_fs_if fs_info = {
639 .sysfs_name = "xgmi_wafl_err_count",
642 if (!adev->gmc.xgmi.supported ||
643 adev->gmc.xgmi.num_physical_nodes == 0)
646 amdgpu_xgmi_reset_ras_error_count(adev);
648 if (!adev->gmc.xgmi.ras_if) {
649 adev->gmc.xgmi.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL);
650 if (!adev->gmc.xgmi.ras_if)
652 adev->gmc.xgmi.ras_if->block = AMDGPU_RAS_BLOCK__XGMI_WAFL;
653 adev->gmc.xgmi.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
654 adev->gmc.xgmi.ras_if->sub_block_index = 0;
655 strcpy(adev->gmc.xgmi.ras_if->name, "xgmi_wafl");
657 ih_info.head = fs_info.head = *adev->gmc.xgmi.ras_if;
658 r = amdgpu_ras_late_init(adev, adev->gmc.xgmi.ras_if,
660 if (r || !amdgpu_ras_is_supported(adev, adev->gmc.xgmi.ras_if->block)) {
661 kfree(adev->gmc.xgmi.ras_if);
662 adev->gmc.xgmi.ras_if = NULL;
668 void amdgpu_xgmi_ras_fini(struct amdgpu_device *adev)
670 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL) &&
671 adev->gmc.xgmi.ras_if) {
672 struct ras_common_if *ras_if = adev->gmc.xgmi.ras_if;
673 struct ras_ih_if ih_info = {
677 amdgpu_ras_late_fini(adev, ras_if, &ih_info);
682 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev,
685 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi;
686 return (addr + xgmi->physical_node_id * xgmi->node_segment_size);
689 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg)
691 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF);
692 WREG32_PCIE(pcs_status_reg, 0);
695 void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev)
699 switch (adev->asic_type) {
701 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++)
702 pcs_clear_status(adev,
703 xgmi_pcs_err_status_reg_arct[i]);
706 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++)
707 pcs_clear_status(adev,
708 xgmi_pcs_err_status_reg_vg20[i]);
715 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev,
725 /* query xgmi pcs error status,
726 * only ue is supported */
727 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_ras_fields); i ++) {
729 xgmi_pcs_ras_fields[i].pcs_err_mask) >>
730 xgmi_pcs_ras_fields[i].pcs_err_shift;
732 dev_info(adev->dev, "%s detected\n",
733 xgmi_pcs_ras_fields[i].err_name);
738 /* query wafl pcs error status,
739 * only ue is supported */
740 for (i = 0; i < ARRAY_SIZE(wafl_pcs_ras_fields); i++) {
742 wafl_pcs_ras_fields[i].pcs_err_mask) >>
743 wafl_pcs_ras_fields[i].pcs_err_shift;
745 dev_info(adev->dev, "%s detected\n",
746 wafl_pcs_ras_fields[i].err_name);
755 int amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev,
756 void *ras_error_status)
758 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
761 uint32_t ue_cnt = 0, ce_cnt = 0;
763 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL))
766 err_data->ue_count = 0;
767 err_data->ce_count = 0;
769 switch (adev->asic_type) {
771 /* check xgmi pcs error */
772 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) {
773 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]);
775 amdgpu_xgmi_query_pcs_error_status(adev,
776 data, &ue_cnt, &ce_cnt, true);
778 /* check wafl pcs error */
779 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) {
780 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]);
782 amdgpu_xgmi_query_pcs_error_status(adev,
783 data, &ue_cnt, &ce_cnt, false);
788 /* check xgmi pcs error */
789 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) {
790 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]);
792 amdgpu_xgmi_query_pcs_error_status(adev,
793 data, &ue_cnt, &ce_cnt, true);
795 /* check wafl pcs error */
796 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) {
797 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]);
799 amdgpu_xgmi_query_pcs_error_status(adev,
800 data, &ue_cnt, &ce_cnt, false);
805 amdgpu_xgmi_reset_ras_error_count(adev);
807 err_data->ue_count += ue_cnt;
808 err_data->ce_count += ce_cnt;