xref: /linux/drivers/gpu/drm/xe/xe_device_sysfs.c (revision 72c181399b01bb4836d1fabaa9f5f6438c82178e)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2023 Intel Corporation
4  */
5 
6 #include <linux/device.h>
7 #include <linux/kobject.h>
8 #include <linux/pci.h>
9 #include <linux/sysfs.h>
10 
11 #include "xe_device.h"
12 #include "xe_device_sysfs.h"
13 #include "xe_mmio.h"
14 #include "xe_pcode_api.h"
15 #include "xe_pcode.h"
16 #include "xe_pm.h"
17 
18 /**
19  * DOC: Xe device sysfs
20  * Xe driver requires exposing certain tunable knobs controlled by user space for
21  * each graphics device. Considering this, we need to add sysfs attributes at device
22  * level granularity.
23  * These sysfs attributes will be available under pci device kobj directory.
24  *
25  * vram_d3cold_threshold - Report/change vram used threshold(in MB) below
26  * which vram save/restore is permissible during runtime D3cold entry/exit.
27  *
28  * lb_fan_control_version - Fan control version provisioned by late binding.
29  * Exposed only if supported by the device.
30  *
31  * lb_voltage_regulator_version - Voltage regulator version provisioned by late
32  * binding. Exposed only if supported by the device.
33  */
34 
35 static ssize_t
vram_d3cold_threshold_show(struct device * dev,struct device_attribute * attr,char * buf)36 vram_d3cold_threshold_show(struct device *dev,
37 			   struct device_attribute *attr, char *buf)
38 {
39 	struct pci_dev *pdev = to_pci_dev(dev);
40 	struct xe_device *xe = pdev_to_xe_device(pdev);
41 	int ret;
42 
43 	xe_pm_runtime_get(xe);
44 	ret = sysfs_emit(buf, "%d\n", xe->d3cold.vram_threshold);
45 	xe_pm_runtime_put(xe);
46 
47 	return ret;
48 }
49 
50 static ssize_t
vram_d3cold_threshold_store(struct device * dev,struct device_attribute * attr,const char * buff,size_t count)51 vram_d3cold_threshold_store(struct device *dev, struct device_attribute *attr,
52 			    const char *buff, size_t count)
53 {
54 	struct pci_dev *pdev = to_pci_dev(dev);
55 	struct xe_device *xe = pdev_to_xe_device(pdev);
56 	u32 vram_d3cold_threshold;
57 	int ret;
58 
59 	ret = kstrtou32(buff, 0, &vram_d3cold_threshold);
60 	if (ret)
61 		return ret;
62 
63 	drm_dbg(&xe->drm, "vram_d3cold_threshold: %u\n", vram_d3cold_threshold);
64 
65 	xe_pm_runtime_get(xe);
66 	ret = xe_pm_set_vram_threshold(xe, vram_d3cold_threshold);
67 	xe_pm_runtime_put(xe);
68 
69 	return ret ?: count;
70 }
71 
72 static DEVICE_ATTR_RW(vram_d3cold_threshold);
73 
74 static struct attribute *vram_attrs[] = {
75 	&dev_attr_vram_d3cold_threshold.attr,
76 	NULL
77 };
78 
79 static const struct attribute_group vram_attr_group = {
80 	.attrs = vram_attrs,
81 };
82 
83 static ssize_t
lb_fan_control_version_show(struct device * dev,struct device_attribute * attr,char * buf)84 lb_fan_control_version_show(struct device *dev, struct device_attribute *attr, char *buf)
85 {
86 	struct xe_device *xe = pdev_to_xe_device(to_pci_dev(dev));
87 	struct xe_tile *root = xe_device_get_root_tile(xe);
88 	u32 cap = 0, ver_low = FAN_TABLE, ver_high = FAN_TABLE;
89 	u16 major = 0, minor = 0, hotfix = 0, build = 0;
90 	int ret;
91 
92 	xe_pm_runtime_get(xe);
93 
94 	ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_CAPABILITY_STATUS, 0),
95 			    &cap, NULL);
96 	if (ret)
97 		goto out;
98 
99 	if (REG_FIELD_GET(V1_FAN_PROVISIONED, cap)) {
100 		ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_VERSION_LOW, 0),
101 				    &ver_low, NULL);
102 		if (ret)
103 			goto out;
104 
105 		ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_VERSION_HIGH, 0),
106 				    &ver_high, NULL);
107 		if (ret)
108 			goto out;
109 
110 		major = REG_FIELD_GET(MAJOR_VERSION_MASK, ver_low);
111 		minor = REG_FIELD_GET(MINOR_VERSION_MASK, ver_low);
112 		hotfix = REG_FIELD_GET(HOTFIX_VERSION_MASK, ver_high);
113 		build = REG_FIELD_GET(BUILD_VERSION_MASK, ver_high);
114 	}
115 out:
116 	xe_pm_runtime_put(xe);
117 
118 	return ret ?: sysfs_emit(buf, "%u.%u.%u.%u\n", major, minor, hotfix, build);
119 }
120 static DEVICE_ATTR_ADMIN_RO(lb_fan_control_version);
121 
122 static ssize_t
lb_voltage_regulator_version_show(struct device * dev,struct device_attribute * attr,char * buf)123 lb_voltage_regulator_version_show(struct device *dev, struct device_attribute *attr, char *buf)
124 {
125 	struct xe_device *xe = pdev_to_xe_device(to_pci_dev(dev));
126 	struct xe_tile *root = xe_device_get_root_tile(xe);
127 	u32 cap = 0, ver_low = VR_CONFIG, ver_high = VR_CONFIG;
128 	u16 major = 0, minor = 0, hotfix = 0, build = 0;
129 	int ret;
130 
131 	xe_pm_runtime_get(xe);
132 
133 	ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_CAPABILITY_STATUS, 0),
134 			    &cap, NULL);
135 	if (ret)
136 		goto out;
137 
138 	if (REG_FIELD_GET(VR_PARAMS_PROVISIONED, cap)) {
139 		ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_VERSION_LOW, 0),
140 				    &ver_low, NULL);
141 		if (ret)
142 			goto out;
143 
144 		ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_VERSION_HIGH, 0),
145 				    &ver_high, NULL);
146 		if (ret)
147 			goto out;
148 
149 		major = REG_FIELD_GET(MAJOR_VERSION_MASK, ver_low);
150 		minor = REG_FIELD_GET(MINOR_VERSION_MASK, ver_low);
151 		hotfix = REG_FIELD_GET(HOTFIX_VERSION_MASK, ver_high);
152 		build = REG_FIELD_GET(BUILD_VERSION_MASK, ver_high);
153 	}
154 out:
155 	xe_pm_runtime_put(xe);
156 
157 	return ret ?: sysfs_emit(buf, "%u.%u.%u.%u\n", major, minor, hotfix, build);
158 }
159 static DEVICE_ATTR_ADMIN_RO(lb_voltage_regulator_version);
160 
161 static struct attribute *late_bind_attrs[] = {
162 	&dev_attr_lb_fan_control_version.attr,
163 	&dev_attr_lb_voltage_regulator_version.attr,
164 	NULL
165 };
166 
late_bind_attr_is_visible(struct kobject * kobj,struct attribute * attr,int n)167 static umode_t late_bind_attr_is_visible(struct kobject *kobj,
168 					 struct attribute *attr, int n)
169 {
170 	struct device *dev = kobj_to_dev(kobj);
171 	struct xe_device *xe = pdev_to_xe_device(to_pci_dev(dev));
172 	struct xe_tile *root = xe_device_get_root_tile(xe);
173 	u32 cap = 0;
174 	int ret;
175 
176 	xe_pm_runtime_get(xe);
177 
178 	ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_CAPABILITY_STATUS, 0),
179 			    &cap, NULL);
180 	xe_pm_runtime_put(xe);
181 	if (ret)
182 		return 0;
183 
184 	if (attr == &dev_attr_lb_fan_control_version.attr &&
185 	    REG_FIELD_GET(V1_FAN_SUPPORTED, cap))
186 		return attr->mode;
187 	if (attr == &dev_attr_lb_voltage_regulator_version.attr &&
188 	    REG_FIELD_GET(VR_PARAMS_SUPPORTED, cap))
189 		return attr->mode;
190 
191 	return 0;
192 }
193 
194 static const struct attribute_group late_bind_attr_group = {
195 	.attrs = late_bind_attrs,
196 	.is_visible = late_bind_attr_is_visible,
197 };
198 
199 /**
200  * DOC: PCIe Gen5 Limitations
201  *
202  * Default link speed of discrete GPUs is determined by configuration parameters
203  * stored in their flash memory, which are subject to override through user
204  * initiated firmware updates. It has been observed that devices configured with
205  * PCIe Gen5 as their default link speed can come across link quality issues due
206  * to host or motherboard limitations and may have to auto-downgrade their link
207  * to PCIe Gen4 speed when faced with unstable link at Gen5, which makes
208  * firmware updates rather risky on such setups. It is required to ensure that
209  * the device is capable of auto-downgrading its link to PCIe Gen4 speed before
210  * pushing the firmware image with PCIe Gen5 as default configuration. This can
211  * be done by reading ``auto_link_downgrade_capable`` sysfs entry, which will
212  * denote if the device is capable of auto-downgrading its link to PCIe Gen4
213  * speed with boolean output value of ``0`` or ``1``, meaning `incapable` or
214  * `capable` respectively.
215  *
216  * .. code-block:: shell
217  *
218  *    $ cat /sys/bus/pci/devices/<bdf>/auto_link_downgrade_capable
219  *
220  * Pushing the firmware image with PCIe Gen5 as default configuration on a auto
221  * link downgrade incapable device and facing link instability due to host or
222  * motherboard limitations can result in driver failing to bind to the device,
223  * making further firmware updates impossible with RMA being the only last
224  * resort.
225  *
226  * Link downgrade status of auto link downgrade capable devices is available
227  * through ``auto_link_downgrade_status`` sysfs entry with boolean output value
228  * of ``0`` or ``1``, where ``0`` means no auto-downgrading was required during
229  * link training (which is the optimal scenario) and ``1`` means the device has
230  * auto-downgraded its link to PCIe Gen4 speed due to unstable Gen5 link.
231  *
232  * .. code-block:: shell
233  *
234  *    $ cat /sys/bus/pci/devices/<bdf>/auto_link_downgrade_status
235  */
236 
237 static ssize_t
auto_link_downgrade_capable_show(struct device * dev,struct device_attribute * attr,char * buf)238 auto_link_downgrade_capable_show(struct device *dev, struct device_attribute *attr, char *buf)
239 {
240 	struct pci_dev *pdev = to_pci_dev(dev);
241 	struct xe_device *xe = pdev_to_xe_device(pdev);
242 	u32 cap, val;
243 
244 	xe_pm_runtime_get(xe);
245 	val = xe_mmio_read32(xe_root_tile_mmio(xe), BMG_PCIE_CAP);
246 	xe_pm_runtime_put(xe);
247 
248 	cap = REG_FIELD_GET(LINK_DOWNGRADE, val);
249 	return sysfs_emit(buf, "%u\n", cap == DOWNGRADE_CAPABLE);
250 }
251 static DEVICE_ATTR_ADMIN_RO(auto_link_downgrade_capable);
252 
253 static ssize_t
auto_link_downgrade_status_show(struct device * dev,struct device_attribute * attr,char * buf)254 auto_link_downgrade_status_show(struct device *dev, struct device_attribute *attr, char *buf)
255 {
256 	struct pci_dev *pdev = to_pci_dev(dev);
257 	struct xe_device *xe = pdev_to_xe_device(pdev);
258 	/* default the auto_link_downgrade status to 0 */
259 	u32 val = 0;
260 	int ret;
261 
262 	xe_pm_runtime_get(xe);
263 	ret = xe_pcode_read(xe_device_get_root_tile(xe),
264 			    PCODE_MBOX(DGFX_PCODE_STATUS, DGFX_GET_INIT_STATUS, 0),
265 			    &val, NULL);
266 	xe_pm_runtime_put(xe);
267 
268 	return ret ?: sysfs_emit(buf, "%u\n", REG_FIELD_GET(DGFX_LINK_DOWNGRADE_STATUS, val));
269 }
270 static DEVICE_ATTR_ADMIN_RO(auto_link_downgrade_status);
271 
272 static struct attribute *auto_link_downgrade_attrs[] = {
273 	&dev_attr_auto_link_downgrade_capable.attr,
274 	&dev_attr_auto_link_downgrade_status.attr,
275 	NULL
276 };
277 
278 static const struct attribute_group auto_link_downgrade_attr_group = {
279 	.attrs = auto_link_downgrade_attrs,
280 };
281 
xe_device_sysfs_init(struct xe_device * xe)282 int xe_device_sysfs_init(struct xe_device *xe)
283 {
284 	struct device *dev = xe->drm.dev;
285 	int ret;
286 
287 	if (xe->d3cold.capable) {
288 		ret = devm_device_add_group(dev, &vram_attr_group);
289 		if (ret)
290 			return ret;
291 	}
292 
293 	if (xe->info.platform == XE_BATTLEMAGE && !IS_SRIOV_VF(xe)) {
294 		ret = devm_device_add_group(dev, &auto_link_downgrade_attr_group);
295 		if (ret)
296 			return ret;
297 
298 		ret = devm_device_add_group(dev, &late_bind_attr_group);
299 		if (ret)
300 			return ret;
301 	}
302 
303 	return 0;
304 }
305