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_ras.h"
29 #include "df/df_3_6_offset.h"
30 #include "xgmi/xgmi_4_0_0_smn.h"
31 #include "xgmi/xgmi_4_0_0_sh_mask.h"
32 #include "wafl/wafl2_4_0_0_smn.h"
33 #include "wafl/wafl2_4_0_0_sh_mask.h"
35 #define smnPCS_XGMI23_PCS_ERROR_STATUS 0x11a01210
36 #define smnPCS_XGMI3X16_PCS_ERROR_STATUS 0x11a0020c
37 #define smnPCS_GOPX1_PCS_ERROR_STATUS 0x12200210
39 static DEFINE_MUTEX(xgmi_mutex);
41 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE 4
43 static LIST_HEAD(xgmi_hive_list);
45 static const int xgmi_pcs_err_status_reg_vg20[] = {
46 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS,
47 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000,
50 static const int wafl_pcs_err_status_reg_vg20[] = {
51 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS,
52 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000,
55 static const int xgmi_pcs_err_status_reg_arct[] = {
56 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS,
57 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000,
58 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000,
59 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000,
60 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000,
61 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000,
65 static const int wafl_pcs_err_status_reg_arct[] = {
66 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS,
67 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000,
70 static const int xgmi23_pcs_err_status_reg_aldebaran[] = {
71 smnPCS_XGMI23_PCS_ERROR_STATUS,
72 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x100000,
73 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x200000,
74 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x300000,
75 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x400000,
76 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x500000,
77 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x600000,
78 smnPCS_XGMI23_PCS_ERROR_STATUS + 0x700000
81 static const int xgmi3x16_pcs_err_status_reg_aldebaran[] = {
82 smnPCS_XGMI3X16_PCS_ERROR_STATUS,
83 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x100000,
84 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x200000,
85 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x300000,
86 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x400000,
87 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x500000,
88 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x600000,
89 smnPCS_XGMI3X16_PCS_ERROR_STATUS + 0x700000
92 static const int walf_pcs_err_status_reg_aldebaran[] = {
93 smnPCS_GOPX1_PCS_ERROR_STATUS,
94 smnPCS_GOPX1_PCS_ERROR_STATUS + 0x100000
97 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = {
98 {"XGMI PCS DataLossErr",
99 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)},
100 {"XGMI PCS TrainingErr",
101 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)},
103 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)},
104 {"XGMI PCS BERExceededErr",
105 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)},
106 {"XGMI PCS TxMetaDataErr",
107 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)},
108 {"XGMI PCS ReplayBufParityErr",
109 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)},
110 {"XGMI PCS DataParityErr",
111 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)},
112 {"XGMI PCS ReplayFifoOverflowErr",
113 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
114 {"XGMI PCS ReplayFifoUnderflowErr",
115 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
116 {"XGMI PCS ElasticFifoOverflowErr",
117 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
118 {"XGMI PCS DeskewErr",
119 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)},
120 {"XGMI PCS DataStartupLimitErr",
121 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)},
122 {"XGMI PCS FCInitTimeoutErr",
123 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)},
124 {"XGMI PCS RecoveryTimeoutErr",
125 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
126 {"XGMI PCS ReadySerialTimeoutErr",
127 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
128 {"XGMI PCS ReadySerialAttemptErr",
129 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
130 {"XGMI PCS RecoveryAttemptErr",
131 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)},
132 {"XGMI PCS RecoveryRelockAttemptErr",
133 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
136 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = {
137 {"WAFL PCS DataLossErr",
138 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)},
139 {"WAFL PCS TrainingErr",
140 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)},
142 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)},
143 {"WAFL PCS BERExceededErr",
144 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)},
145 {"WAFL PCS TxMetaDataErr",
146 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)},
147 {"WAFL PCS ReplayBufParityErr",
148 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)},
149 {"WAFL PCS DataParityErr",
150 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)},
151 {"WAFL PCS ReplayFifoOverflowErr",
152 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)},
153 {"WAFL PCS ReplayFifoUnderflowErr",
154 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)},
155 {"WAFL PCS ElasticFifoOverflowErr",
156 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)},
157 {"WAFL PCS DeskewErr",
158 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)},
159 {"WAFL PCS DataStartupLimitErr",
160 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)},
161 {"WAFL PCS FCInitTimeoutErr",
162 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)},
163 {"WAFL PCS RecoveryTimeoutErr",
164 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)},
165 {"WAFL PCS ReadySerialTimeoutErr",
166 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)},
167 {"WAFL PCS ReadySerialAttemptErr",
168 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)},
169 {"WAFL PCS RecoveryAttemptErr",
170 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)},
171 {"WAFL PCS RecoveryRelockAttemptErr",
172 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)},
176 * DOC: AMDGPU XGMI Support
178 * XGMI is a high speed interconnect that joins multiple GPU cards
179 * into a homogeneous memory space that is organized by a collective
180 * hive ID and individual node IDs, both of which are 64-bit numbers.
182 * The file xgmi_device_id contains the unique per GPU device ID and
183 * is stored in the /sys/class/drm/card${cardno}/device/ directory.
185 * Inside the device directory a sub-directory 'xgmi_hive_info' is
186 * created which contains the hive ID and the list of nodes.
188 * The hive ID is stored in:
189 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id
191 * The node information is stored in numbered directories:
192 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id
194 * Each device has their own xgmi_hive_info direction with a mirror
195 * set of node sub-directories.
197 * The XGMI memory space is built by contiguously adding the power of
198 * two padded VRAM space from each node to each other.
202 static struct attribute amdgpu_xgmi_hive_id = {
203 .name = "xgmi_hive_id",
207 static struct attribute *amdgpu_xgmi_hive_attrs[] = {
208 &amdgpu_xgmi_hive_id,
212 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj,
213 struct attribute *attr, char *buf)
215 struct amdgpu_hive_info *hive = container_of(
216 kobj, struct amdgpu_hive_info, kobj);
218 if (attr == &amdgpu_xgmi_hive_id)
219 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id);
224 static void amdgpu_xgmi_hive_release(struct kobject *kobj)
226 struct amdgpu_hive_info *hive = container_of(
227 kobj, struct amdgpu_hive_info, kobj);
229 mutex_destroy(&hive->hive_lock);
233 static const struct sysfs_ops amdgpu_xgmi_hive_ops = {
234 .show = amdgpu_xgmi_show_attrs,
237 struct kobj_type amdgpu_xgmi_hive_type = {
238 .release = amdgpu_xgmi_hive_release,
239 .sysfs_ops = &amdgpu_xgmi_hive_ops,
240 .default_attrs = amdgpu_xgmi_hive_attrs,
243 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev,
244 struct device_attribute *attr,
247 struct drm_device *ddev = dev_get_drvdata(dev);
248 struct amdgpu_device *adev = drm_to_adev(ddev);
250 return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id);
254 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801)
255 static ssize_t amdgpu_xgmi_show_error(struct device *dev,
256 struct device_attribute *attr,
259 struct drm_device *ddev = dev_get_drvdata(dev);
260 struct amdgpu_device *adev = drm_to_adev(ddev);
261 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in;
263 unsigned int error_count = 0;
265 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200);
266 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208);
268 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in);
269 if (fica_out != 0x1f)
270 pr_err("xGMI error counters not enabled!\n");
272 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in);
274 if ((fica_out & 0xffff) == 2)
275 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63);
277 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0);
279 return sysfs_emit(buf, "%u\n", error_count);
283 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL);
284 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL);
286 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev,
287 struct amdgpu_hive_info *hive)
290 char node[10] = { 0 };
292 /* Create xgmi device id file */
293 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id);
295 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n");
299 /* Create xgmi error file */
300 ret = device_create_file(adev->dev, &dev_attr_xgmi_error);
302 pr_err("failed to create xgmi_error\n");
305 /* Create sysfs link to hive info folder on the first device */
306 if (hive->kobj.parent != (&adev->dev->kobj)) {
307 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj,
310 dev_err(adev->dev, "XGMI: Failed to create link to hive info");
315 sprintf(node, "node%d", atomic_read(&hive->number_devices));
316 /* Create sysfs link form the hive folder to yourself */
317 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node);
319 dev_err(adev->dev, "XGMI: Failed to create link from hive info");
327 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique);
330 device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
336 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev,
337 struct amdgpu_hive_info *hive)
340 memset(node, 0, sizeof(node));
342 device_remove_file(adev->dev, &dev_attr_xgmi_device_id);
343 device_remove_file(adev->dev, &dev_attr_xgmi_error);
345 if (hive->kobj.parent != (&adev->dev->kobj))
346 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info");
348 sprintf(node, "node%d", atomic_read(&hive->number_devices));
349 sysfs_remove_link(&hive->kobj, node);
355 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev)
357 struct amdgpu_hive_info *hive = NULL;
360 if (!adev->gmc.xgmi.hive_id)
364 kobject_get(&adev->hive->kobj);
368 mutex_lock(&xgmi_mutex);
370 list_for_each_entry(hive, &xgmi_hive_list, node) {
371 if (hive->hive_id == adev->gmc.xgmi.hive_id)
375 hive = kzalloc(sizeof(*hive), GFP_KERNEL);
377 dev_err(adev->dev, "XGMI: allocation failed\n");
382 /* initialize new hive if not exist */
383 ret = kobject_init_and_add(&hive->kobj,
384 &amdgpu_xgmi_hive_type,
386 "%s", "xgmi_hive_info");
388 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n");
394 hive->hive_id = adev->gmc.xgmi.hive_id;
395 INIT_LIST_HEAD(&hive->device_list);
396 INIT_LIST_HEAD(&hive->node);
397 mutex_init(&hive->hive_lock);
398 atomic_set(&hive->in_reset, 0);
399 atomic_set(&hive->number_devices, 0);
400 task_barrier_init(&hive->tb);
401 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN;
402 hive->hi_req_gpu = NULL;
404 * hive pstate on boot is high in vega20 so we have to go to low
405 * pstate on after boot.
407 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE;
408 list_add_tail(&hive->node, &xgmi_hive_list);
412 kobject_get(&hive->kobj);
413 mutex_unlock(&xgmi_mutex);
417 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive)
420 kobject_put(&hive->kobj);
423 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate)
426 struct amdgpu_hive_info *hive;
427 struct amdgpu_device *request_adev;
428 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20;
431 hive = amdgpu_get_xgmi_hive(adev);
435 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev;
436 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN;
437 amdgpu_put_xgmi_hive(hive);
438 /* fw bug so temporarily disable pstate switching */
441 if (!hive || adev->asic_type != CHIP_VEGA20)
444 mutex_lock(&hive->hive_lock);
447 hive->hi_req_count++;
449 hive->hi_req_count--;
452 * Vega20 only needs single peer to request pstate high for the hive to
453 * go high but all peers must request pstate low for the hive to go low
455 if (hive->pstate == pstate ||
456 (!is_hi_req && hive->hi_req_count && !init_low))
459 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate);
461 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate);
463 dev_err(request_adev->dev,
464 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d",
465 request_adev->gmc.xgmi.node_id,
466 request_adev->gmc.xgmi.hive_id, ret);
471 hive->pstate = hive->hi_req_count ?
472 hive->pstate : AMDGPU_XGMI_PSTATE_MIN;
474 hive->pstate = pstate;
475 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ?
479 mutex_unlock(&hive->hive_lock);
483 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev)
487 /* Each psp need to set the latest topology */
488 ret = psp_xgmi_set_topology_info(&adev->psp,
489 atomic_read(&hive->number_devices),
490 &adev->psp.xgmi_context.top_info);
493 "XGMI: Set topology failure on device %llx, hive %llx, ret %d",
494 adev->gmc.xgmi.node_id,
495 adev->gmc.xgmi.hive_id, ret);
502 * NOTE psp_xgmi_node_info.num_hops layout is as follows:
503 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved)
504 * num_hops[5:3] = reserved
505 * num_hops[2:0] = number of hops
507 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev,
508 struct amdgpu_device *peer_adev)
510 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
511 uint8_t num_hops_mask = 0x7;
514 for (i = 0 ; i < top->num_nodes; ++i)
515 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
516 return top->nodes[i].num_hops & num_hops_mask;
520 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev,
521 struct amdgpu_device *peer_adev)
523 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info;
526 for (i = 0 ; i < top->num_nodes; ++i)
527 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id)
528 return top->nodes[i].num_links;
533 * Devices that support extended data require the entire hive to initialize with
534 * the shared memory buffer flag set.
536 * Hive locks and conditions apply - see amdgpu_xgmi_add_device
538 static int amdgpu_xgmi_initialize_hive_get_data_partition(struct amdgpu_hive_info *hive,
539 bool set_extended_data)
541 struct amdgpu_device *tmp_adev;
544 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
545 ret = psp_xgmi_initialize(&tmp_adev->psp, set_extended_data, false);
547 dev_err(tmp_adev->dev,
548 "XGMI: Failed to initialize xgmi session for data partition %i\n",
558 int amdgpu_xgmi_add_device(struct amdgpu_device *adev)
560 struct psp_xgmi_topology_info *top_info;
561 struct amdgpu_hive_info *hive;
562 struct amdgpu_xgmi *entry;
563 struct amdgpu_device *tmp_adev = NULL;
565 int count = 0, ret = 0;
567 if (!adev->gmc.xgmi.supported)
570 if (!adev->gmc.xgmi.pending_reset &&
571 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
572 ret = psp_xgmi_initialize(&adev->psp, false, true);
575 "XGMI: Failed to initialize xgmi session\n");
579 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id);
582 "XGMI: Failed to get hive id\n");
586 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id);
589 "XGMI: Failed to get node id\n");
593 adev->gmc.xgmi.hive_id = 16;
594 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16;
597 hive = amdgpu_get_xgmi_hive(adev);
601 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n",
602 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id);
605 mutex_lock(&hive->hive_lock);
607 top_info = &adev->psp.xgmi_context.top_info;
609 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list);
610 list_for_each_entry(entry, &hive->device_list, head)
611 top_info->nodes[count++].node_id = entry->node_id;
612 top_info->num_nodes = count;
613 atomic_set(&hive->number_devices, count);
615 task_barrier_add_task(&hive->tb);
617 if (!adev->gmc.xgmi.pending_reset &&
618 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) {
619 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
620 /* update node list for other device in the hive */
621 if (tmp_adev != adev) {
622 top_info = &tmp_adev->psp.xgmi_context.top_info;
623 top_info->nodes[count - 1].node_id =
624 adev->gmc.xgmi.node_id;
625 top_info->num_nodes = count;
627 ret = amdgpu_xgmi_update_topology(hive, tmp_adev);
632 /* get latest topology info for each device from psp */
633 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
634 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
635 &tmp_adev->psp.xgmi_context.top_info, false);
637 dev_err(tmp_adev->dev,
638 "XGMI: Get topology failure on device %llx, hive %llx, ret %d",
639 tmp_adev->gmc.xgmi.node_id,
640 tmp_adev->gmc.xgmi.hive_id, ret);
641 /* To do : continue with some node failed or disable the whole hive */
646 /* get topology again for hives that support extended data */
647 if (adev->psp.xgmi_context.supports_extended_data) {
649 /* initialize the hive to get extended data. */
650 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, true);
654 /* get the extended data. */
655 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
656 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count,
657 &tmp_adev->psp.xgmi_context.top_info, true);
659 dev_err(tmp_adev->dev,
660 "XGMI: Get topology for extended data failure on device %llx, hive %llx, ret %d",
661 tmp_adev->gmc.xgmi.node_id,
662 tmp_adev->gmc.xgmi.hive_id, ret);
667 /* initialize the hive to get non-extended data for the next round. */
668 ret = amdgpu_xgmi_initialize_hive_get_data_partition(hive, false);
675 if (!ret && !adev->gmc.xgmi.pending_reset)
676 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive);
679 mutex_unlock(&hive->hive_lock);
683 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n",
684 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id);
686 amdgpu_put_xgmi_hive(hive);
687 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n",
688 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id,
695 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev)
697 struct amdgpu_hive_info *hive = adev->hive;
699 if (!adev->gmc.xgmi.supported)
705 mutex_lock(&hive->hive_lock);
706 task_barrier_rem_task(&hive->tb);
707 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive);
708 if (hive->hi_req_gpu == adev)
709 hive->hi_req_gpu = NULL;
710 list_del(&adev->gmc.xgmi.head);
711 mutex_unlock(&hive->hive_lock);
713 amdgpu_put_xgmi_hive(hive);
716 if (atomic_dec_return(&hive->number_devices) == 0) {
717 /* Remove the hive from global hive list */
718 mutex_lock(&xgmi_mutex);
719 list_del(&hive->node);
720 mutex_unlock(&xgmi_mutex);
722 amdgpu_put_xgmi_hive(hive);
725 return psp_xgmi_terminate(&adev->psp);
728 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev)
731 struct ras_ih_if ih_info = {
734 struct ras_fs_if fs_info = {
735 .sysfs_name = "xgmi_wafl_err_count",
738 if (!adev->gmc.xgmi.supported ||
739 adev->gmc.xgmi.num_physical_nodes == 0)
742 adev->gmc.xgmi.ras_funcs->reset_ras_error_count(adev);
744 if (!adev->gmc.xgmi.ras_if) {
745 adev->gmc.xgmi.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL);
746 if (!adev->gmc.xgmi.ras_if)
748 adev->gmc.xgmi.ras_if->block = AMDGPU_RAS_BLOCK__XGMI_WAFL;
749 adev->gmc.xgmi.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
750 adev->gmc.xgmi.ras_if->sub_block_index = 0;
752 ih_info.head = fs_info.head = *adev->gmc.xgmi.ras_if;
753 r = amdgpu_ras_late_init(adev, adev->gmc.xgmi.ras_if,
755 if (r || !amdgpu_ras_is_supported(adev, adev->gmc.xgmi.ras_if->block)) {
756 kfree(adev->gmc.xgmi.ras_if);
757 adev->gmc.xgmi.ras_if = NULL;
763 static void amdgpu_xgmi_ras_fini(struct amdgpu_device *adev)
765 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL) &&
766 adev->gmc.xgmi.ras_if) {
767 struct ras_common_if *ras_if = adev->gmc.xgmi.ras_if;
768 struct ras_ih_if ih_info = {
772 amdgpu_ras_late_fini(adev, ras_if, &ih_info);
777 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev,
780 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi;
781 return (addr + xgmi->physical_node_id * xgmi->node_segment_size);
784 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg)
786 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF);
787 WREG32_PCIE(pcs_status_reg, 0);
790 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev)
794 switch (adev->asic_type) {
796 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++)
797 pcs_clear_status(adev,
798 xgmi_pcs_err_status_reg_arct[i]);
801 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++)
802 pcs_clear_status(adev,
803 xgmi_pcs_err_status_reg_vg20[i]);
806 for (i = 0; i < ARRAY_SIZE(xgmi23_pcs_err_status_reg_aldebaran); i++)
807 pcs_clear_status(adev,
808 xgmi23_pcs_err_status_reg_aldebaran[i]);
809 for (i = 0; i < ARRAY_SIZE(xgmi23_pcs_err_status_reg_aldebaran); i++)
810 pcs_clear_status(adev,
811 xgmi23_pcs_err_status_reg_aldebaran[i]);
812 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++)
813 pcs_clear_status(adev,
814 walf_pcs_err_status_reg_aldebaran[i]);
821 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev,
831 /* query xgmi pcs error status,
832 * only ue is supported */
833 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_ras_fields); i ++) {
835 xgmi_pcs_ras_fields[i].pcs_err_mask) >>
836 xgmi_pcs_ras_fields[i].pcs_err_shift;
838 dev_info(adev->dev, "%s detected\n",
839 xgmi_pcs_ras_fields[i].err_name);
844 /* query wafl pcs error status,
845 * only ue is supported */
846 for (i = 0; i < ARRAY_SIZE(wafl_pcs_ras_fields); i++) {
848 wafl_pcs_ras_fields[i].pcs_err_mask) >>
849 wafl_pcs_ras_fields[i].pcs_err_shift;
851 dev_info(adev->dev, "%s detected\n",
852 wafl_pcs_ras_fields[i].err_name);
861 static int amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev,
862 void *ras_error_status)
864 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status;
867 uint32_t ue_cnt = 0, ce_cnt = 0;
869 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL))
872 err_data->ue_count = 0;
873 err_data->ce_count = 0;
875 switch (adev->asic_type) {
877 /* check xgmi pcs error */
878 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) {
879 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]);
881 amdgpu_xgmi_query_pcs_error_status(adev,
882 data, &ue_cnt, &ce_cnt, true);
884 /* check wafl pcs error */
885 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) {
886 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]);
888 amdgpu_xgmi_query_pcs_error_status(adev,
889 data, &ue_cnt, &ce_cnt, false);
893 /* check xgmi pcs error */
894 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) {
895 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]);
897 amdgpu_xgmi_query_pcs_error_status(adev,
898 data, &ue_cnt, &ce_cnt, true);
900 /* check wafl pcs error */
901 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) {
902 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]);
904 amdgpu_xgmi_query_pcs_error_status(adev,
905 data, &ue_cnt, &ce_cnt, false);
909 /* check xgmi23 pcs error */
910 for (i = 0; i < ARRAY_SIZE(xgmi23_pcs_err_status_reg_aldebaran); i++) {
911 data = RREG32_PCIE(xgmi23_pcs_err_status_reg_aldebaran[i]);
913 amdgpu_xgmi_query_pcs_error_status(adev,
914 data, &ue_cnt, &ce_cnt, true);
916 /* check xgmi3x16 pcs error */
917 for (i = 0; i < ARRAY_SIZE(xgmi3x16_pcs_err_status_reg_aldebaran); i++) {
918 data = RREG32_PCIE(xgmi3x16_pcs_err_status_reg_aldebaran[i]);
920 amdgpu_xgmi_query_pcs_error_status(adev,
921 data, &ue_cnt, &ce_cnt, true);
923 /* check wafl pcs error */
924 for (i = 0; i < ARRAY_SIZE(walf_pcs_err_status_reg_aldebaran); i++) {
925 data = RREG32_PCIE(walf_pcs_err_status_reg_aldebaran[i]);
927 amdgpu_xgmi_query_pcs_error_status(adev,
928 data, &ue_cnt, &ce_cnt, false);
932 dev_warn(adev->dev, "XGMI RAS error query not supported");
936 adev->gmc.xgmi.ras_funcs->reset_ras_error_count(adev);
938 err_data->ue_count += ue_cnt;
939 err_data->ce_count += ce_cnt;
944 const struct amdgpu_xgmi_ras_funcs xgmi_ras_funcs = {
945 .ras_late_init = amdgpu_xgmi_ras_late_init,
946 .ras_fini = amdgpu_xgmi_ras_fini,
947 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count,
948 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count,