1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2023 Intel Corporation
4 */
5
6 #include <drm/drm_managed.h>
7
8 #include "xe_force_wake.h"
9 #include "xe_device.h"
10 #include "xe_gt.h"
11 #include "xe_gt_idle.h"
12 #include "xe_gt_sysfs.h"
13 #include "xe_guc_pc.h"
14 #include "regs/xe_gt_regs.h"
15 #include "xe_macros.h"
16 #include "xe_mmio.h"
17 #include "xe_pm.h"
18 #include "xe_sriov.h"
19
20 /**
21 * DOC: Xe GT Idle
22 *
23 * Contains functions that init GT idle features like C6
24 *
25 * device/gt#/gtidle/name - name of the state
26 * device/gt#/gtidle/idle_residency_ms - Provides residency of the idle state in ms
27 * device/gt#/gtidle/idle_status - Provides current idle state
28 */
29
dev_to_gtidle(struct device * dev)30 static struct xe_gt_idle *dev_to_gtidle(struct device *dev)
31 {
32 struct kobject *kobj = &dev->kobj;
33
34 return &kobj_to_gt(kobj->parent)->gtidle;
35 }
36
gtidle_to_gt(struct xe_gt_idle * gtidle)37 static struct xe_gt *gtidle_to_gt(struct xe_gt_idle *gtidle)
38 {
39 return container_of(gtidle, struct xe_gt, gtidle);
40 }
41
gtidle_to_pc(struct xe_gt_idle * gtidle)42 static struct xe_guc_pc *gtidle_to_pc(struct xe_gt_idle *gtidle)
43 {
44 return >idle_to_gt(gtidle)->uc.guc.pc;
45 }
46
47 static struct xe_device *
pc_to_xe(struct xe_guc_pc * pc)48 pc_to_xe(struct xe_guc_pc *pc)
49 {
50 struct xe_guc *guc = container_of(pc, struct xe_guc, pc);
51 struct xe_gt *gt = container_of(guc, struct xe_gt, uc.guc);
52
53 return gt_to_xe(gt);
54 }
55
gt_idle_state_to_string(enum xe_gt_idle_state state)56 static const char *gt_idle_state_to_string(enum xe_gt_idle_state state)
57 {
58 switch (state) {
59 case GT_IDLE_C0:
60 return "gt-c0";
61 case GT_IDLE_C6:
62 return "gt-c6";
63 default:
64 return "unknown";
65 }
66 }
67
get_residency_ms(struct xe_gt_idle * gtidle,u64 cur_residency)68 static u64 get_residency_ms(struct xe_gt_idle *gtidle, u64 cur_residency)
69 {
70 u64 delta, overflow_residency, prev_residency;
71
72 lockdep_assert_held(>idle->lock);
73
74 overflow_residency = BIT_ULL(32);
75
76 /*
77 * Counter wrap handling
78 * Store previous hw counter values for counter wrap-around handling
79 * Relying on sufficient frequency of queries otherwise counters can still wrap.
80 */
81 prev_residency = gtidle->prev_residency;
82 gtidle->prev_residency = cur_residency;
83
84 /* delta */
85 if (cur_residency >= prev_residency)
86 delta = cur_residency - prev_residency;
87 else
88 delta = cur_residency + (overflow_residency - prev_residency);
89
90 /* Add delta to extended raw driver copy of idle residency */
91 cur_residency = gtidle->cur_residency + delta;
92 gtidle->cur_residency = cur_residency;
93
94 /* residency multiplier in ns, convert to ms */
95 cur_residency = mul_u64_u32_div(cur_residency, gtidle->residency_multiplier, 1e6);
96
97 return cur_residency;
98 }
99
xe_gt_idle_enable_pg(struct xe_gt * gt)100 void xe_gt_idle_enable_pg(struct xe_gt *gt)
101 {
102 struct xe_device *xe = gt_to_xe(gt);
103 struct xe_gt_idle *gtidle = >->gtidle;
104 struct xe_mmio *mmio = >->mmio;
105 u32 vcs_mask, vecs_mask;
106 unsigned int fw_ref;
107 int i, j;
108
109 if (IS_SRIOV_VF(xe))
110 return;
111
112 /* Disable CPG for PVC */
113 if (xe->info.platform == XE_PVC)
114 return;
115
116 xe_device_assert_mem_access(gt_to_xe(gt));
117
118 vcs_mask = xe_hw_engine_mask_per_class(gt, XE_ENGINE_CLASS_VIDEO_DECODE);
119 vecs_mask = xe_hw_engine_mask_per_class(gt, XE_ENGINE_CLASS_VIDEO_ENHANCE);
120
121 if (vcs_mask || vecs_mask)
122 gtidle->powergate_enable = MEDIA_POWERGATE_ENABLE;
123
124 if (!xe_gt_is_media_type(gt))
125 gtidle->powergate_enable |= RENDER_POWERGATE_ENABLE;
126
127 if (xe->info.platform != XE_DG1) {
128 for (i = XE_HW_ENGINE_VCS0, j = 0; i <= XE_HW_ENGINE_VCS7; ++i, ++j) {
129 if ((gt->info.engine_mask & BIT(i)))
130 gtidle->powergate_enable |= (VDN_HCP_POWERGATE_ENABLE(j) |
131 VDN_MFXVDENC_POWERGATE_ENABLE(j));
132 }
133 }
134
135 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
136 if (xe->info.skip_guc_pc) {
137 /*
138 * GuC sets the hysteresis value when GuC PC is enabled
139 * else set it to 25 (25 * 1.28us)
140 */
141 xe_mmio_write32(mmio, MEDIA_POWERGATE_IDLE_HYSTERESIS, 25);
142 xe_mmio_write32(mmio, RENDER_POWERGATE_IDLE_HYSTERESIS, 25);
143 }
144
145 xe_mmio_write32(mmio, POWERGATE_ENABLE, gtidle->powergate_enable);
146 xe_force_wake_put(gt_to_fw(gt), fw_ref);
147 }
148
xe_gt_idle_disable_pg(struct xe_gt * gt)149 void xe_gt_idle_disable_pg(struct xe_gt *gt)
150 {
151 struct xe_gt_idle *gtidle = >->gtidle;
152 unsigned int fw_ref;
153
154 if (IS_SRIOV_VF(gt_to_xe(gt)))
155 return;
156
157 xe_device_assert_mem_access(gt_to_xe(gt));
158 gtidle->powergate_enable = 0;
159
160 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
161 xe_mmio_write32(>->mmio, POWERGATE_ENABLE, gtidle->powergate_enable);
162 xe_force_wake_put(gt_to_fw(gt), fw_ref);
163 }
164
165 /**
166 * xe_gt_idle_pg_print - Xe powergating info
167 * @gt: GT object
168 * @p: drm_printer.
169 *
170 * This function prints the powergating information
171 *
172 * Return: 0 on success, negative error code otherwise
173 */
xe_gt_idle_pg_print(struct xe_gt * gt,struct drm_printer * p)174 int xe_gt_idle_pg_print(struct xe_gt *gt, struct drm_printer *p)
175 {
176 struct xe_gt_idle *gtidle = >->gtidle;
177 struct xe_device *xe = gt_to_xe(gt);
178 enum xe_gt_idle_state state;
179 u32 pg_enabled, pg_status = 0;
180 u32 vcs_mask, vecs_mask;
181 unsigned int fw_ref;
182 int n;
183 /*
184 * Media Slices
185 *
186 * Slice 0: VCS0, VCS1, VECS0
187 * Slice 1: VCS2, VCS3, VECS1
188 * Slice 2: VCS4, VCS5, VECS2
189 * Slice 3: VCS6, VCS7, VECS3
190 */
191 static const struct {
192 u64 engines;
193 u32 status_bit;
194 } media_slices[] = {
195 {(BIT(XE_HW_ENGINE_VCS0) | BIT(XE_HW_ENGINE_VCS1) |
196 BIT(XE_HW_ENGINE_VECS0)), MEDIA_SLICE0_AWAKE_STATUS},
197
198 {(BIT(XE_HW_ENGINE_VCS2) | BIT(XE_HW_ENGINE_VCS3) |
199 BIT(XE_HW_ENGINE_VECS1)), MEDIA_SLICE1_AWAKE_STATUS},
200
201 {(BIT(XE_HW_ENGINE_VCS4) | BIT(XE_HW_ENGINE_VCS5) |
202 BIT(XE_HW_ENGINE_VECS2)), MEDIA_SLICE2_AWAKE_STATUS},
203
204 {(BIT(XE_HW_ENGINE_VCS6) | BIT(XE_HW_ENGINE_VCS7) |
205 BIT(XE_HW_ENGINE_VECS3)), MEDIA_SLICE3_AWAKE_STATUS},
206 };
207
208 if (xe->info.platform == XE_PVC) {
209 drm_printf(p, "Power Gating not supported\n");
210 return 0;
211 }
212
213 state = gtidle->idle_status(gtidle_to_pc(gtidle));
214 pg_enabled = gtidle->powergate_enable;
215
216 /* Do not wake the GT to read powergating status */
217 if (state != GT_IDLE_C6) {
218 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
219 if (!fw_ref)
220 return -ETIMEDOUT;
221
222 pg_enabled = xe_mmio_read32(>->mmio, POWERGATE_ENABLE);
223 pg_status = xe_mmio_read32(>->mmio, POWERGATE_DOMAIN_STATUS);
224
225 xe_force_wake_put(gt_to_fw(gt), fw_ref);
226 }
227
228 if (gt->info.engine_mask & XE_HW_ENGINE_RCS_MASK) {
229 drm_printf(p, "Render Power Gating Enabled: %s\n",
230 str_yes_no(pg_enabled & RENDER_POWERGATE_ENABLE));
231
232 drm_printf(p, "Render Power Gate Status: %s\n",
233 str_up_down(pg_status & RENDER_AWAKE_STATUS));
234 }
235
236 vcs_mask = xe_hw_engine_mask_per_class(gt, XE_ENGINE_CLASS_VIDEO_DECODE);
237 vecs_mask = xe_hw_engine_mask_per_class(gt, XE_ENGINE_CLASS_VIDEO_ENHANCE);
238
239 /* Print media CPG status only if media is present */
240 if (vcs_mask || vecs_mask) {
241 drm_printf(p, "Media Power Gating Enabled: %s\n",
242 str_yes_no(pg_enabled & MEDIA_POWERGATE_ENABLE));
243
244 for (n = 0; n < ARRAY_SIZE(media_slices); n++)
245 if (gt->info.engine_mask & media_slices[n].engines)
246 drm_printf(p, "Media Slice%d Power Gate Status: %s\n", n,
247 str_up_down(pg_status & media_slices[n].status_bit));
248 }
249 return 0;
250 }
251
name_show(struct device * dev,struct device_attribute * attr,char * buff)252 static ssize_t name_show(struct device *dev,
253 struct device_attribute *attr, char *buff)
254 {
255 struct xe_gt_idle *gtidle = dev_to_gtidle(dev);
256 struct xe_guc_pc *pc = gtidle_to_pc(gtidle);
257 ssize_t ret;
258
259 xe_pm_runtime_get(pc_to_xe(pc));
260 ret = sysfs_emit(buff, "%s\n", gtidle->name);
261 xe_pm_runtime_put(pc_to_xe(pc));
262
263 return ret;
264 }
265 static DEVICE_ATTR_RO(name);
266
idle_status_show(struct device * dev,struct device_attribute * attr,char * buff)267 static ssize_t idle_status_show(struct device *dev,
268 struct device_attribute *attr, char *buff)
269 {
270 struct xe_gt_idle *gtidle = dev_to_gtidle(dev);
271 struct xe_guc_pc *pc = gtidle_to_pc(gtidle);
272 enum xe_gt_idle_state state;
273
274 xe_pm_runtime_get(pc_to_xe(pc));
275 state = gtidle->idle_status(pc);
276 xe_pm_runtime_put(pc_to_xe(pc));
277
278 return sysfs_emit(buff, "%s\n", gt_idle_state_to_string(state));
279 }
280
xe_gt_idle_residency_msec(struct xe_gt_idle * gtidle)281 u64 xe_gt_idle_residency_msec(struct xe_gt_idle *gtidle)
282 {
283 struct xe_guc_pc *pc = gtidle_to_pc(gtidle);
284 u64 residency;
285 unsigned long flags;
286
287 raw_spin_lock_irqsave(>idle->lock, flags);
288 residency = get_residency_ms(gtidle, gtidle->idle_residency(pc));
289 raw_spin_unlock_irqrestore(>idle->lock, flags);
290
291 return residency;
292 }
293
294 static DEVICE_ATTR_RO(idle_status);
idle_residency_ms_show(struct device * dev,struct device_attribute * attr,char * buff)295 static ssize_t idle_residency_ms_show(struct device *dev,
296 struct device_attribute *attr, char *buff)
297 {
298 struct xe_gt_idle *gtidle = dev_to_gtidle(dev);
299 struct xe_guc_pc *pc = gtidle_to_pc(gtidle);
300 u64 residency;
301
302 xe_pm_runtime_get(pc_to_xe(pc));
303 residency = xe_gt_idle_residency_msec(gtidle);
304 xe_pm_runtime_put(pc_to_xe(pc));
305
306 return sysfs_emit(buff, "%llu\n", residency);
307 }
308 static DEVICE_ATTR_RO(idle_residency_ms);
309
310 static const struct attribute *gt_idle_attrs[] = {
311 &dev_attr_name.attr,
312 &dev_attr_idle_status.attr,
313 &dev_attr_idle_residency_ms.attr,
314 NULL,
315 };
316
gt_idle_fini(void * arg)317 static void gt_idle_fini(void *arg)
318 {
319 struct kobject *kobj = arg;
320 struct xe_gt *gt = kobj_to_gt(kobj->parent);
321 unsigned int fw_ref;
322
323 xe_gt_idle_disable_pg(gt);
324
325 if (gt_to_xe(gt)->info.skip_guc_pc) {
326 fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FW_GT);
327 xe_gt_idle_disable_c6(gt);
328 xe_force_wake_put(gt_to_fw(gt), fw_ref);
329 }
330
331 sysfs_remove_files(kobj, gt_idle_attrs);
332 kobject_put(kobj);
333 }
334
xe_gt_idle_init(struct xe_gt_idle * gtidle)335 int xe_gt_idle_init(struct xe_gt_idle *gtidle)
336 {
337 struct xe_gt *gt = gtidle_to_gt(gtidle);
338 struct xe_device *xe = gt_to_xe(gt);
339 struct kobject *kobj;
340 int err;
341
342 if (IS_SRIOV_VF(xe))
343 return 0;
344
345 kobj = kobject_create_and_add("gtidle", gt->sysfs);
346 if (!kobj)
347 return -ENOMEM;
348
349 raw_spin_lock_init(>idle->lock);
350
351 if (xe_gt_is_media_type(gt)) {
352 snprintf(gtidle->name, sizeof(gtidle->name), "gt%d-mc", gt->info.id);
353 gtidle->idle_residency = xe_guc_pc_mc6_residency;
354 } else {
355 snprintf(gtidle->name, sizeof(gtidle->name), "gt%d-rc", gt->info.id);
356 gtidle->idle_residency = xe_guc_pc_rc6_residency;
357 }
358
359 /* Multiplier for Residency counter in units of 1.28us */
360 gtidle->residency_multiplier = 1280;
361 gtidle->idle_status = xe_guc_pc_c_status;
362
363 err = sysfs_create_files(kobj, gt_idle_attrs);
364 if (err) {
365 kobject_put(kobj);
366 return err;
367 }
368
369 xe_gt_idle_enable_pg(gt);
370
371 return devm_add_action_or_reset(xe->drm.dev, gt_idle_fini, kobj);
372 }
373
xe_gt_idle_enable_c6(struct xe_gt * gt)374 void xe_gt_idle_enable_c6(struct xe_gt *gt)
375 {
376 xe_device_assert_mem_access(gt_to_xe(gt));
377 xe_force_wake_assert_held(gt_to_fw(gt), XE_FW_GT);
378
379 if (IS_SRIOV_VF(gt_to_xe(gt)))
380 return;
381
382 /* Units of 1280 ns for a total of 5s */
383 xe_mmio_write32(>->mmio, RC_IDLE_HYSTERSIS, 0x3B9ACA);
384 /* Enable RC6 */
385 xe_mmio_write32(>->mmio, RC_CONTROL,
386 RC_CTL_HW_ENABLE | RC_CTL_TO_MODE | RC_CTL_RC6_ENABLE);
387 }
388
xe_gt_idle_disable_c6(struct xe_gt * gt)389 void xe_gt_idle_disable_c6(struct xe_gt *gt)
390 {
391 xe_device_assert_mem_access(gt_to_xe(gt));
392 xe_force_wake_assert_held(gt_to_fw(gt), XE_FW_GT);
393
394 if (IS_SRIOV_VF(gt_to_xe(gt)))
395 return;
396
397 xe_mmio_write32(>->mmio, RC_CONTROL, 0);
398 xe_mmio_write32(>->mmio, RC_STATE, 0);
399 }
400