OSDN Git Service

47c6dbe8fff3e892190ef3417f4b258b90a2dc5f
[uclinux-h8/linux.git] / drivers / gpu / drm / nouveau / nvkm / subdev / pmu / gk20a.c
1 /*
2  * Copyright (c) 2014, NVIDIA CORPORATION. All rights reserved.
3  *
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:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
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 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
19  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
20  * DEALINGS IN THE SOFTWARE.
21  */
22 #include "priv.h"
23
24 #include <subdev/clk.h>
25 #include <subdev/timer.h>
26 #include <subdev/volt.h>
27
28 #define BUSY_SLOT       0
29 #define CLK_SLOT        7
30
31 struct gk20a_pmu_dvfs_data {
32         int p_load_target;
33         int p_load_max;
34         int p_smooth;
35         unsigned int avg_load;
36 };
37
38 struct gk20a_pmu {
39         struct nvkm_pmu base;
40         struct nvkm_alarm alarm;
41         struct gk20a_pmu_dvfs_data *data;
42 };
43
44 struct gk20a_pmu_dvfs_dev_status {
45         unsigned long total;
46         unsigned long busy;
47         int cur_state;
48 };
49
50 static int
51 gk20a_pmu_dvfs_target(struct gk20a_pmu *pmu, int *state)
52 {
53         struct nvkm_clk *clk = nvkm_clk(pmu);
54
55         return nvkm_clk_astate(clk, *state, 0, false);
56 }
57
58 static int
59 gk20a_pmu_dvfs_get_cur_state(struct gk20a_pmu *pmu, int *state)
60 {
61         struct nvkm_clk *clk = nvkm_clk(pmu);
62
63         *state = clk->pstate;
64         return 0;
65 }
66
67 static int
68 gk20a_pmu_dvfs_get_target_state(struct gk20a_pmu *pmu,
69                                 int *state, int load)
70 {
71         struct gk20a_pmu_dvfs_data *data = pmu->data;
72         struct nvkm_clk *clk = nvkm_clk(pmu);
73         int cur_level, level;
74
75         /* For GK20A, the performance level is directly mapped to pstate */
76         level = cur_level = clk->pstate;
77
78         if (load > data->p_load_max) {
79                 level = min(clk->state_nr - 1, level + (clk->state_nr / 3));
80         } else {
81                 level += ((load - data->p_load_target) * 10 /
82                                 data->p_load_target) / 2;
83                 level = max(0, level);
84                 level = min(clk->state_nr - 1, level);
85         }
86
87         nvkm_trace(&pmu->base.subdev, "cur level = %d, new level = %d\n",
88                    cur_level, level);
89
90         *state = level;
91
92         if (level == cur_level)
93                 return 0;
94         else
95                 return 1;
96 }
97
98 static int
99 gk20a_pmu_dvfs_get_dev_status(struct gk20a_pmu *pmu,
100                               struct gk20a_pmu_dvfs_dev_status *status)
101 {
102         struct nvkm_device *device = pmu->base.subdev.device;
103         status->busy = nvkm_rd32(device, 0x10a508 + (BUSY_SLOT * 0x10));
104         status->total= nvkm_rd32(device, 0x10a508 + (CLK_SLOT * 0x10));
105         return 0;
106 }
107
108 static void
109 gk20a_pmu_dvfs_reset_dev_status(struct gk20a_pmu *pmu)
110 {
111         struct nvkm_device *device = pmu->base.subdev.device;
112         nvkm_wr32(device, 0x10a508 + (BUSY_SLOT * 0x10), 0x80000000);
113         nvkm_wr32(device, 0x10a508 + (CLK_SLOT * 0x10), 0x80000000);
114 }
115
116 static void
117 gk20a_pmu_dvfs_work(struct nvkm_alarm *alarm)
118 {
119         struct gk20a_pmu *pmu =
120                 container_of(alarm, struct gk20a_pmu, alarm);
121         struct gk20a_pmu_dvfs_data *data = pmu->data;
122         struct gk20a_pmu_dvfs_dev_status status;
123         struct nvkm_subdev *subdev = &pmu->base.subdev;
124         struct nvkm_device *device = subdev->device;
125         struct nvkm_clk *clk = device->clk;
126         struct nvkm_volt *volt = device->volt;
127         u32 utilization = 0;
128         int state, ret;
129
130         /*
131          * The PMU is initialized before CLK and VOLT, so we have to make sure the
132          * CLK and VOLT are ready here.
133          */
134         if (!clk || !volt)
135                 goto resched;
136
137         ret = gk20a_pmu_dvfs_get_dev_status(pmu, &status);
138         if (ret) {
139                 nvkm_warn(subdev, "failed to get device status\n");
140                 goto resched;
141         }
142
143         if (status.total)
144                 utilization = div_u64((u64)status.busy * 100, status.total);
145
146         data->avg_load = (data->p_smooth * data->avg_load) + utilization;
147         data->avg_load /= data->p_smooth + 1;
148         nvkm_trace(subdev, "utilization = %d %%, avg_load = %d %%\n",
149                    utilization, data->avg_load);
150
151         ret = gk20a_pmu_dvfs_get_cur_state(pmu, &state);
152         if (ret) {
153                 nvkm_warn(subdev, "failed to get current state\n");
154                 goto resched;
155         }
156
157         if (gk20a_pmu_dvfs_get_target_state(pmu, &state, data->avg_load)) {
158                 nvkm_trace(subdev, "set new state to %d\n", state);
159                 gk20a_pmu_dvfs_target(pmu, &state);
160         }
161
162 resched:
163         gk20a_pmu_dvfs_reset_dev_status(pmu);
164         nvkm_timer_alarm(pmu, 100000000, alarm);
165 }
166
167 static int
168 gk20a_pmu_fini(struct nvkm_object *object, bool suspend)
169 {
170         struct gk20a_pmu *pmu = (void *)object;
171
172         nvkm_timer_alarm_cancel(pmu, &pmu->alarm);
173
174         return nvkm_subdev_fini_old(&pmu->base.subdev, suspend);
175 }
176
177 static int
178 gk20a_pmu_init(struct nvkm_object *object)
179 {
180         struct gk20a_pmu *pmu = (void *)object;
181         struct nvkm_device *device = pmu->base.subdev.device;
182         int ret;
183
184         ret = nvkm_subdev_init_old(&pmu->base.subdev);
185         if (ret)
186                 return ret;
187
188         pmu->base.pgob = nvkm_pmu_pgob;
189
190         /* init pwr perf counter */
191         nvkm_wr32(device, 0x10a504 + (BUSY_SLOT * 0x10), 0x00200001);
192         nvkm_wr32(device, 0x10a50c + (BUSY_SLOT * 0x10), 0x00000002);
193         nvkm_wr32(device, 0x10a50c + (CLK_SLOT * 0x10), 0x00000003);
194
195         nvkm_timer_alarm(pmu, 2000000000, &pmu->alarm);
196         return ret;
197 }
198
199 static struct gk20a_pmu_dvfs_data
200 gk20a_dvfs_data= {
201         .p_load_target = 70,
202         .p_load_max = 90,
203         .p_smooth = 1,
204 };
205
206 static int
207 gk20a_pmu_ctor(struct nvkm_object *parent, struct nvkm_object *engine,
208                struct nvkm_oclass *oclass, void *data, u32 size,
209                struct nvkm_object **pobject)
210 {
211         struct gk20a_pmu *pmu;
212         int ret;
213
214         ret = nvkm_pmu_create(parent, engine, oclass, &pmu);
215         *pobject = nv_object(pmu);
216         if (ret)
217                 return ret;
218
219         pmu->data = &gk20a_dvfs_data;
220
221         nvkm_alarm_init(&pmu->alarm, gk20a_pmu_dvfs_work);
222         return 0;
223 }
224
225 struct nvkm_oclass *
226 gk20a_pmu_oclass = &(struct nvkm_pmu_impl) {
227         .base.handle = NV_SUBDEV(PMU, 0xea),
228         .base.ofuncs = &(struct nvkm_ofuncs) {
229                 .ctor = gk20a_pmu_ctor,
230                 .dtor = _nvkm_pmu_dtor,
231                 .init = gk20a_pmu_init,
232                 .fini = gk20a_pmu_fini,
233         },
234 }.base;