xref: /linux/drivers/gpu/drm/xe/xe_device_sysfs.c (revision c0d6f52f9b62479d61f8cd4faf9fb2f8bce6e301)
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
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 
42 	return sysfs_emit(buf, "%d\n", xe->d3cold.vram_threshold);
43 }
44 
45 static ssize_t
46 vram_d3cold_threshold_store(struct device *dev, struct device_attribute *attr,
47 			    const char *buff, size_t count)
48 {
49 	struct pci_dev *pdev = to_pci_dev(dev);
50 	struct xe_device *xe = pdev_to_xe_device(pdev);
51 	u32 vram_d3cold_threshold;
52 	int ret;
53 
54 	ret = kstrtou32(buff, 0, &vram_d3cold_threshold);
55 	if (ret)
56 		return ret;
57 
58 	drm_dbg(&xe->drm, "vram_d3cold_threshold: %u\n", vram_d3cold_threshold);
59 
60 	guard(xe_pm_runtime)(xe);
61 	ret = xe_pm_set_vram_threshold(xe, vram_d3cold_threshold);
62 
63 	return ret ?: count;
64 }
65 
66 static DEVICE_ATTR_RW(vram_d3cold_threshold);
67 
68 static struct attribute *vram_attrs[] = {
69 	&dev_attr_vram_d3cold_threshold.attr,
70 	NULL
71 };
72 
73 static const struct attribute_group vram_attr_group = {
74 	.attrs = vram_attrs,
75 };
76 
77 static ssize_t
78 lb_fan_control_version_show(struct device *dev, struct device_attribute *attr, char *buf)
79 {
80 	struct xe_device *xe = pdev_to_xe_device(to_pci_dev(dev));
81 	struct xe_tile *root = xe_device_get_root_tile(xe);
82 	u32 cap = 0, ver_low = FAN_TABLE, ver_high = FAN_TABLE;
83 	u16 major = 0, minor = 0, hotfix = 0, build = 0;
84 	int ret;
85 
86 	guard(xe_pm_runtime)(xe);
87 
88 	ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_CAPABILITY_STATUS, 0),
89 			    &cap, NULL);
90 	if (ret)
91 		return ret;
92 
93 	if (REG_FIELD_GET(V1_FAN_PROVISIONED, cap)) {
94 		ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_VERSION_LOW, 0),
95 				    &ver_low, NULL);
96 		if (ret)
97 			return ret;
98 
99 		ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_VERSION_HIGH, 0),
100 				    &ver_high, NULL);
101 		if (ret)
102 			return ret;
103 
104 		major = REG_FIELD_GET(MAJOR_VERSION_MASK, ver_low);
105 		minor = REG_FIELD_GET(MINOR_VERSION_MASK, ver_low);
106 		hotfix = REG_FIELD_GET(HOTFIX_VERSION_MASK, ver_high);
107 		build = REG_FIELD_GET(BUILD_VERSION_MASK, ver_high);
108 	}
109 
110 	return sysfs_emit(buf, "%u.%u.%u.%u\n", major, minor, hotfix, build);
111 }
112 static DEVICE_ATTR_ADMIN_RO(lb_fan_control_version);
113 
114 static ssize_t
115 lb_voltage_regulator_version_show(struct device *dev, struct device_attribute *attr, char *buf)
116 {
117 	struct xe_device *xe = pdev_to_xe_device(to_pci_dev(dev));
118 	struct xe_tile *root = xe_device_get_root_tile(xe);
119 	u32 cap = 0, ver_low = VR_CONFIG, ver_high = VR_CONFIG;
120 	u16 major = 0, minor = 0, hotfix = 0, build = 0;
121 	int ret;
122 
123 	guard(xe_pm_runtime)(xe);
124 
125 	ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_CAPABILITY_STATUS, 0),
126 			    &cap, NULL);
127 	if (ret)
128 		return ret;
129 
130 	if (REG_FIELD_GET(VR_PARAMS_PROVISIONED, cap)) {
131 		ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_VERSION_LOW, 0),
132 				    &ver_low, NULL);
133 		if (ret)
134 			return ret;
135 
136 		ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_VERSION_HIGH, 0),
137 				    &ver_high, NULL);
138 		if (ret)
139 			return ret;
140 
141 		major = REG_FIELD_GET(MAJOR_VERSION_MASK, ver_low);
142 		minor = REG_FIELD_GET(MINOR_VERSION_MASK, ver_low);
143 		hotfix = REG_FIELD_GET(HOTFIX_VERSION_MASK, ver_high);
144 		build = REG_FIELD_GET(BUILD_VERSION_MASK, ver_high);
145 	}
146 
147 	return sysfs_emit(buf, "%u.%u.%u.%u\n", major, minor, hotfix, build);
148 }
149 static DEVICE_ATTR_ADMIN_RO(lb_voltage_regulator_version);
150 
151 static struct attribute *late_bind_attrs[] = {
152 	&dev_attr_lb_fan_control_version.attr,
153 	&dev_attr_lb_voltage_regulator_version.attr,
154 	NULL
155 };
156 
157 static umode_t late_bind_attr_is_visible(struct kobject *kobj,
158 					 struct attribute *attr, int n)
159 {
160 	struct device *dev = kobj_to_dev(kobj);
161 	struct xe_device *xe = pdev_to_xe_device(to_pci_dev(dev));
162 	struct xe_tile *root = xe_device_get_root_tile(xe);
163 	u32 cap = 0;
164 	int ret;
165 
166 	ret = xe_pcode_read(root, PCODE_MBOX(PCODE_LATE_BINDING, GET_CAPABILITY_STATUS, 0),
167 			    &cap, NULL);
168 	if (ret)
169 		return 0;
170 
171 	if (attr == &dev_attr_lb_fan_control_version.attr &&
172 	    REG_FIELD_GET(V1_FAN_SUPPORTED, cap))
173 		return attr->mode;
174 	if (attr == &dev_attr_lb_voltage_regulator_version.attr &&
175 	    REG_FIELD_GET(VR_PARAMS_SUPPORTED, cap))
176 		return attr->mode;
177 
178 	return 0;
179 }
180 
181 static const struct attribute_group late_bind_attr_group = {
182 	.attrs = late_bind_attrs,
183 	.is_visible = late_bind_attr_is_visible,
184 };
185 
186 /**
187  * DOC: PCIe Gen5 Limitations
188  *
189  * Default link speed of discrete GPUs is determined by configuration parameters
190  * stored in their flash memory, which are subject to override through user
191  * initiated firmware updates. It has been observed that devices configured with
192  * PCIe Gen5 as their default link speed can come across link quality issues due
193  * to host or motherboard limitations and may have to auto-downgrade their link
194  * to PCIe Gen4 speed when faced with unstable link at Gen5, which makes
195  * firmware updates rather risky on such setups. It is required to ensure that
196  * the device is capable of auto-downgrading its link to PCIe Gen4 speed before
197  * pushing the firmware image with PCIe Gen5 as default configuration. This can
198  * be done by reading ``auto_link_downgrade_capable`` sysfs entry, which will
199  * denote if the device is capable of auto-downgrading its link to PCIe Gen4
200  * speed with boolean output value of ``0`` or ``1``, meaning `incapable` or
201  * `capable` respectively.
202  *
203  * .. code-block:: shell
204  *
205  *    $ cat /sys/bus/pci/devices/<bdf>/auto_link_downgrade_capable
206  *
207  * Pushing the firmware image with PCIe Gen5 as default configuration on a auto
208  * link downgrade incapable device and facing link instability due to host or
209  * motherboard limitations can result in driver failing to bind to the device,
210  * making further firmware updates impossible with RMA being the only last
211  * resort.
212  *
213  * Link downgrade status of auto link downgrade capable devices is available
214  * through ``auto_link_downgrade_status`` sysfs entry with boolean output value
215  * of ``0`` or ``1``, where ``0`` means no auto-downgrading was required during
216  * link training (which is the optimal scenario) and ``1`` means the device has
217  * auto-downgraded its link to PCIe Gen4 speed due to unstable Gen5 link.
218  *
219  * .. code-block:: shell
220  *
221  *    $ cat /sys/bus/pci/devices/<bdf>/auto_link_downgrade_status
222  */
223 
224 static ssize_t
225 auto_link_downgrade_capable_show(struct device *dev, struct device_attribute *attr, char *buf)
226 {
227 	struct pci_dev *pdev = to_pci_dev(dev);
228 	struct xe_device *xe = pdev_to_xe_device(pdev);
229 	u32 cap, val;
230 
231 	guard(xe_pm_runtime)(xe);
232 	val = xe_mmio_read32(xe_root_tile_mmio(xe), BMG_PCIE_CAP);
233 
234 	cap = REG_FIELD_GET(LINK_DOWNGRADE, val);
235 	return sysfs_emit(buf, "%u\n", cap == DOWNGRADE_CAPABLE);
236 }
237 static DEVICE_ATTR_ADMIN_RO(auto_link_downgrade_capable);
238 
239 static ssize_t
240 auto_link_downgrade_status_show(struct device *dev, struct device_attribute *attr, char *buf)
241 {
242 	struct pci_dev *pdev = to_pci_dev(dev);
243 	struct xe_device *xe = pdev_to_xe_device(pdev);
244 	/* default the auto_link_downgrade status to 0 */
245 	u32 val = 0;
246 	int ret;
247 
248 	guard(xe_pm_runtime)(xe);
249 	ret = xe_pcode_read(xe_device_get_root_tile(xe),
250 			    PCODE_MBOX(DGFX_PCODE_STATUS, DGFX_GET_INIT_STATUS, 0),
251 			    &val, NULL);
252 
253 	return ret ?: sysfs_emit(buf, "%u\n", REG_FIELD_GET(DGFX_LINK_DOWNGRADE_STATUS, val));
254 }
255 static DEVICE_ATTR_ADMIN_RO(auto_link_downgrade_status);
256 
257 static struct attribute *auto_link_downgrade_attrs[] = {
258 	&dev_attr_auto_link_downgrade_capable.attr,
259 	&dev_attr_auto_link_downgrade_status.attr,
260 	NULL
261 };
262 
263 static const struct attribute_group auto_link_downgrade_attr_group = {
264 	.attrs = auto_link_downgrade_attrs,
265 };
266 
267 int xe_device_sysfs_init(struct xe_device *xe)
268 {
269 	struct device *dev = xe->drm.dev;
270 	int ret;
271 
272 	if (xe->d3cold.capable) {
273 		ret = devm_device_add_group(dev, &vram_attr_group);
274 		if (ret)
275 			return ret;
276 	}
277 
278 	if (xe->info.platform == XE_BATTLEMAGE && !IS_SRIOV_VF(xe)) {
279 		ret = devm_device_add_group(dev, &auto_link_downgrade_attr_group);
280 		if (ret)
281 			return ret;
282 
283 		ret = devm_device_add_group(dev, &late_bind_attr_group);
284 		if (ret)
285 			return ret;
286 	}
287 
288 	return 0;
289 }
290