1 // SPDX-License-Identifier: GPL-2.0 or MIT 2 /* Copyright 2019 Collabora ltd. */ 3 4 #include <linux/clk.h> 5 #include <linux/devfreq.h> 6 #include <linux/devfreq_cooling.h> 7 #include <linux/platform_device.h> 8 #include <linux/pm_opp.h> 9 10 #include <drm/drm_managed.h> 11 12 #include "panthor_devfreq.h" 13 #include "panthor_device.h" 14 15 /** 16 * struct panthor_devfreq - Device frequency management 17 */ 18 struct panthor_devfreq { 19 /** @devfreq: devfreq device. */ 20 struct devfreq *devfreq; 21 22 /** @gov_data: Governor data. */ 23 struct devfreq_simple_ondemand_data gov_data; 24 25 /** @busy_time: Busy time. */ 26 ktime_t busy_time; 27 28 /** @idle_time: Idle time. */ 29 ktime_t idle_time; 30 31 /** @time_last_update: Last update time. */ 32 ktime_t time_last_update; 33 34 /** @last_busy_state: True if the GPU was busy last time we updated the state. */ 35 bool last_busy_state; 36 37 /** 38 * @lock: Lock used to protect busy_time, idle_time, time_last_update and 39 * last_busy_state. 40 * 41 * These fields can be accessed concurrently by panthor_devfreq_get_dev_status() 42 * and panthor_devfreq_record_{busy,idle}(). 43 */ 44 spinlock_t lock; 45 }; 46 47 static void panthor_devfreq_update_utilization(struct panthor_devfreq *pdevfreq) 48 { 49 ktime_t now, last; 50 51 now = ktime_get(); 52 last = pdevfreq->time_last_update; 53 54 if (pdevfreq->last_busy_state) 55 pdevfreq->busy_time += ktime_sub(now, last); 56 else 57 pdevfreq->idle_time += ktime_sub(now, last); 58 59 pdevfreq->time_last_update = now; 60 } 61 62 static int panthor_devfreq_target(struct device *dev, unsigned long *freq, 63 u32 flags) 64 { 65 struct dev_pm_opp *opp; 66 67 opp = devfreq_recommended_opp(dev, freq, flags); 68 if (IS_ERR(opp)) 69 return PTR_ERR(opp); 70 dev_pm_opp_put(opp); 71 72 return dev_pm_opp_set_rate(dev, *freq); 73 } 74 75 static void panthor_devfreq_reset(struct panthor_devfreq *pdevfreq) 76 { 77 pdevfreq->busy_time = 0; 78 pdevfreq->idle_time = 0; 79 pdevfreq->time_last_update = ktime_get(); 80 } 81 82 static int panthor_devfreq_get_dev_status(struct device *dev, 83 struct devfreq_dev_status *status) 84 { 85 struct panthor_device *ptdev = dev_get_drvdata(dev); 86 struct panthor_devfreq *pdevfreq = ptdev->devfreq; 87 unsigned long irqflags; 88 89 status->current_frequency = clk_get_rate(ptdev->clks.core); 90 91 spin_lock_irqsave(&pdevfreq->lock, irqflags); 92 93 panthor_devfreq_update_utilization(pdevfreq); 94 95 status->total_time = ktime_to_ns(ktime_add(pdevfreq->busy_time, 96 pdevfreq->idle_time)); 97 98 status->busy_time = ktime_to_ns(pdevfreq->busy_time); 99 100 panthor_devfreq_reset(pdevfreq); 101 102 spin_unlock_irqrestore(&pdevfreq->lock, irqflags); 103 104 drm_dbg(&ptdev->base, "busy %lu total %lu %lu %% freq %lu MHz\n", 105 status->busy_time, status->total_time, 106 status->busy_time / (status->total_time / 100), 107 status->current_frequency / 1000 / 1000); 108 109 return 0; 110 } 111 112 static struct devfreq_dev_profile panthor_devfreq_profile = { 113 .timer = DEVFREQ_TIMER_DELAYED, 114 .polling_ms = 50, /* ~3 frames */ 115 .target = panthor_devfreq_target, 116 .get_dev_status = panthor_devfreq_get_dev_status, 117 }; 118 119 int panthor_devfreq_init(struct panthor_device *ptdev) 120 { 121 /* There's actually 2 regulators (mali and sram), but the OPP core only 122 * supports one. 123 * 124 * We assume the sram regulator is coupled with the mali one and let 125 * the coupling logic deal with voltage updates. 126 */ 127 static const char * const reg_names[] = { "mali", NULL }; 128 struct thermal_cooling_device *cooling; 129 struct device *dev = ptdev->base.dev; 130 struct panthor_devfreq *pdevfreq; 131 struct dev_pm_opp *opp; 132 unsigned long cur_freq; 133 int ret; 134 135 pdevfreq = drmm_kzalloc(&ptdev->base, sizeof(*ptdev->devfreq), GFP_KERNEL); 136 if (!pdevfreq) 137 return -ENOMEM; 138 139 ptdev->devfreq = pdevfreq; 140 141 ret = devm_pm_opp_set_regulators(dev, reg_names); 142 if (ret) { 143 if (ret != -EPROBE_DEFER) 144 DRM_DEV_ERROR(dev, "Couldn't set OPP regulators\n"); 145 146 return ret; 147 } 148 149 ret = devm_pm_opp_of_add_table(dev); 150 if (ret) 151 return ret; 152 153 spin_lock_init(&pdevfreq->lock); 154 155 panthor_devfreq_reset(pdevfreq); 156 157 cur_freq = clk_get_rate(ptdev->clks.core); 158 159 opp = devfreq_recommended_opp(dev, &cur_freq, 0); 160 if (IS_ERR(opp)) 161 return PTR_ERR(opp); 162 163 panthor_devfreq_profile.initial_freq = cur_freq; 164 165 /* Regulator coupling only takes care of synchronizing/balancing voltage 166 * updates, but the coupled regulator needs to be enabled manually. 167 * 168 * We use devm_regulator_get_enable_optional() and keep the sram supply 169 * enabled until the device is removed, just like we do for the mali 170 * supply, which is enabled when dev_pm_opp_set_opp(dev, opp) is called, 171 * and disabled when the opp_table is torn down, using the devm action. 172 * 173 * If we really care about disabling regulators on suspend, we should: 174 * - use devm_regulator_get_optional() here 175 * - call dev_pm_opp_set_opp(dev, NULL) before leaving this function 176 * (this disables the regulator passed to the OPP layer) 177 * - call dev_pm_opp_set_opp(dev, NULL) and 178 * regulator_disable(ptdev->regulators.sram) in 179 * panthor_devfreq_suspend() 180 * - call dev_pm_opp_set_opp(dev, default_opp) and 181 * regulator_enable(ptdev->regulators.sram) in 182 * panthor_devfreq_resume() 183 * 184 * But without knowing if it's beneficial or not (in term of power 185 * consumption), or how much it slows down the suspend/resume steps, 186 * let's just keep regulators enabled for the device lifetime. 187 */ 188 ret = devm_regulator_get_enable_optional(dev, "sram"); 189 if (ret && ret != -ENODEV) { 190 if (ret != -EPROBE_DEFER) 191 DRM_DEV_ERROR(dev, "Couldn't retrieve/enable sram supply\n"); 192 return ret; 193 } 194 195 /* 196 * Set the recommend OPP this will enable and configure the regulator 197 * if any and will avoid a switch off by regulator_late_cleanup() 198 */ 199 ret = dev_pm_opp_set_opp(dev, opp); 200 if (ret) { 201 DRM_DEV_ERROR(dev, "Couldn't set recommended OPP\n"); 202 return ret; 203 } 204 205 dev_pm_opp_put(opp); 206 207 /* 208 * Setup default thresholds for the simple_ondemand governor. 209 * The values are chosen based on experiments. 210 */ 211 pdevfreq->gov_data.upthreshold = 45; 212 pdevfreq->gov_data.downdifferential = 5; 213 214 pdevfreq->devfreq = devm_devfreq_add_device(dev, &panthor_devfreq_profile, 215 DEVFREQ_GOV_SIMPLE_ONDEMAND, 216 &pdevfreq->gov_data); 217 if (IS_ERR(pdevfreq->devfreq)) { 218 DRM_DEV_ERROR(dev, "Couldn't initialize GPU devfreq\n"); 219 ret = PTR_ERR(pdevfreq->devfreq); 220 pdevfreq->devfreq = NULL; 221 return ret; 222 } 223 224 cooling = devfreq_cooling_em_register(pdevfreq->devfreq, NULL); 225 if (IS_ERR(cooling)) 226 DRM_DEV_INFO(dev, "Failed to register cooling device\n"); 227 228 return 0; 229 } 230 231 int panthor_devfreq_resume(struct panthor_device *ptdev) 232 { 233 struct panthor_devfreq *pdevfreq = ptdev->devfreq; 234 235 if (!pdevfreq->devfreq) 236 return 0; 237 238 panthor_devfreq_reset(pdevfreq); 239 240 return devfreq_resume_device(pdevfreq->devfreq); 241 } 242 243 int panthor_devfreq_suspend(struct panthor_device *ptdev) 244 { 245 struct panthor_devfreq *pdevfreq = ptdev->devfreq; 246 247 if (!pdevfreq->devfreq) 248 return 0; 249 250 return devfreq_suspend_device(pdevfreq->devfreq); 251 } 252 253 void panthor_devfreq_record_busy(struct panthor_device *ptdev) 254 { 255 struct panthor_devfreq *pdevfreq = ptdev->devfreq; 256 unsigned long irqflags; 257 258 if (!pdevfreq->devfreq) 259 return; 260 261 spin_lock_irqsave(&pdevfreq->lock, irqflags); 262 263 panthor_devfreq_update_utilization(pdevfreq); 264 pdevfreq->last_busy_state = true; 265 266 spin_unlock_irqrestore(&pdevfreq->lock, irqflags); 267 } 268 269 void panthor_devfreq_record_idle(struct panthor_device *ptdev) 270 { 271 struct panthor_devfreq *pdevfreq = ptdev->devfreq; 272 unsigned long irqflags; 273 274 if (!pdevfreq->devfreq) 275 return; 276 277 spin_lock_irqsave(&pdevfreq->lock, irqflags); 278 279 panthor_devfreq_update_utilization(pdevfreq); 280 pdevfreq->last_busy_state = false; 281 282 spin_unlock_irqrestore(&pdevfreq->lock, irqflags); 283 } 284