1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  cpuidle-pseries - idle state cpuidle driver.
4  *  Adapted from drivers/idle/intel_idle.c and
5  *  drivers/acpi/processor_idle.c
6  *
7  */
8 
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 #include <linux/init.h>
12 #include <linux/moduleparam.h>
13 #include <linux/cpuidle.h>
14 #include <linux/cpu.h>
15 #include <linux/notifier.h>
16 
17 #include <asm/paca.h>
18 #include <asm/reg.h>
19 #include <asm/machdep.h>
20 #include <asm/firmware.h>
21 #include <asm/runlatch.h>
22 #include <asm/plpar_wrappers.h>
23 
24 struct cpuidle_driver pseries_idle_driver = {
25 	.name             = "pseries_idle",
26 	.owner            = THIS_MODULE,
27 };
28 
29 static int max_idle_state __read_mostly;
30 static struct cpuidle_state *cpuidle_state_table __read_mostly;
31 static u64 snooze_timeout __read_mostly;
32 static bool snooze_timeout_en __read_mostly;
33 
34 static inline void idle_loop_prolog(unsigned long *in_purr)
35 {
36 	ppc64_runlatch_off();
37 	*in_purr = mfspr(SPRN_PURR);
38 	/*
39 	 * Indicate to the HV that we are idle. Now would be
40 	 * a good time to find other work to dispatch.
41 	 */
42 	get_lppaca()->idle = 1;
43 }
44 
45 static inline void idle_loop_epilog(unsigned long in_purr)
46 {
47 	u64 wait_cycles;
48 
49 	wait_cycles = be64_to_cpu(get_lppaca()->wait_state_cycles);
50 	wait_cycles += mfspr(SPRN_PURR) - in_purr;
51 	get_lppaca()->wait_state_cycles = cpu_to_be64(wait_cycles);
52 	get_lppaca()->idle = 0;
53 
54 	ppc64_runlatch_on();
55 }
56 
57 static int snooze_loop(struct cpuidle_device *dev,
58 			struct cpuidle_driver *drv,
59 			int index)
60 {
61 	unsigned long in_purr;
62 	u64 snooze_exit_time;
63 
64 	set_thread_flag(TIF_POLLING_NRFLAG);
65 
66 	idle_loop_prolog(&in_purr);
67 	local_irq_enable();
68 	snooze_exit_time = get_tb() + snooze_timeout;
69 
70 	while (!need_resched()) {
71 		HMT_low();
72 		HMT_very_low();
73 		if (likely(snooze_timeout_en) && get_tb() > snooze_exit_time) {
74 			/*
75 			 * Task has not woken up but we are exiting the polling
76 			 * loop anyway. Require a barrier after polling is
77 			 * cleared to order subsequent test of need_resched().
78 			 */
79 			clear_thread_flag(TIF_POLLING_NRFLAG);
80 			smp_mb();
81 			break;
82 		}
83 	}
84 
85 	HMT_medium();
86 	clear_thread_flag(TIF_POLLING_NRFLAG);
87 
88 	local_irq_disable();
89 
90 	idle_loop_epilog(in_purr);
91 
92 	return index;
93 }
94 
95 static void check_and_cede_processor(void)
96 {
97 	/*
98 	 * Ensure our interrupt state is properly tracked,
99 	 * also checks if no interrupt has occurred while we
100 	 * were soft-disabled
101 	 */
102 	if (prep_irq_for_idle()) {
103 		cede_processor();
104 #ifdef CONFIG_TRACE_IRQFLAGS
105 		/* Ensure that H_CEDE returns with IRQs on */
106 		if (WARN_ON(!(mfmsr() & MSR_EE)))
107 			__hard_irq_enable();
108 #endif
109 	}
110 }
111 
112 static int dedicated_cede_loop(struct cpuidle_device *dev,
113 				struct cpuidle_driver *drv,
114 				int index)
115 {
116 	unsigned long in_purr;
117 
118 	idle_loop_prolog(&in_purr);
119 	get_lppaca()->donate_dedicated_cpu = 1;
120 
121 	HMT_medium();
122 	check_and_cede_processor();
123 
124 	local_irq_disable();
125 	get_lppaca()->donate_dedicated_cpu = 0;
126 
127 	idle_loop_epilog(in_purr);
128 
129 	return index;
130 }
131 
132 static int shared_cede_loop(struct cpuidle_device *dev,
133 			struct cpuidle_driver *drv,
134 			int index)
135 {
136 	unsigned long in_purr;
137 
138 	idle_loop_prolog(&in_purr);
139 
140 	/*
141 	 * Yield the processor to the hypervisor.  We return if
142 	 * an external interrupt occurs (which are driven prior
143 	 * to returning here) or if a prod occurs from another
144 	 * processor. When returning here, external interrupts
145 	 * are enabled.
146 	 */
147 	check_and_cede_processor();
148 
149 	local_irq_disable();
150 	idle_loop_epilog(in_purr);
151 
152 	return index;
153 }
154 
155 /*
156  * States for dedicated partition case.
157  */
158 static struct cpuidle_state dedicated_states[] = {
159 	{ /* Snooze */
160 		.name = "snooze",
161 		.desc = "snooze",
162 		.exit_latency = 0,
163 		.target_residency = 0,
164 		.enter = &snooze_loop },
165 	{ /* CEDE */
166 		.name = "CEDE",
167 		.desc = "CEDE",
168 		.exit_latency = 10,
169 		.target_residency = 100,
170 		.enter = &dedicated_cede_loop },
171 };
172 
173 /*
174  * States for shared partition case.
175  */
176 static struct cpuidle_state shared_states[] = {
177 	{ /* Snooze */
178 		.name = "snooze",
179 		.desc = "snooze",
180 		.exit_latency = 0,
181 		.target_residency = 0,
182 		.enter = &snooze_loop },
183 	{ /* Shared Cede */
184 		.name = "Shared Cede",
185 		.desc = "Shared Cede",
186 		.exit_latency = 10,
187 		.target_residency = 100,
188 		.enter = &shared_cede_loop },
189 };
190 
191 static int pseries_cpuidle_cpu_online(unsigned int cpu)
192 {
193 	struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
194 
195 	if (dev && cpuidle_get_driver()) {
196 		cpuidle_pause_and_lock();
197 		cpuidle_enable_device(dev);
198 		cpuidle_resume_and_unlock();
199 	}
200 	return 0;
201 }
202 
203 static int pseries_cpuidle_cpu_dead(unsigned int cpu)
204 {
205 	struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
206 
207 	if (dev && cpuidle_get_driver()) {
208 		cpuidle_pause_and_lock();
209 		cpuidle_disable_device(dev);
210 		cpuidle_resume_and_unlock();
211 	}
212 	return 0;
213 }
214 
215 /*
216  * pseries_cpuidle_driver_init()
217  */
218 static int pseries_cpuidle_driver_init(void)
219 {
220 	int idle_state;
221 	struct cpuidle_driver *drv = &pseries_idle_driver;
222 
223 	drv->state_count = 0;
224 
225 	for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
226 		/* Is the state not enabled? */
227 		if (cpuidle_state_table[idle_state].enter == NULL)
228 			continue;
229 
230 		drv->states[drv->state_count] =	/* structure copy */
231 			cpuidle_state_table[idle_state];
232 
233 		drv->state_count += 1;
234 	}
235 
236 	return 0;
237 }
238 
239 /*
240  * pseries_idle_probe()
241  * Choose state table for shared versus dedicated partition
242  */
243 static int pseries_idle_probe(void)
244 {
245 
246 	if (cpuidle_disable != IDLE_NO_OVERRIDE)
247 		return -ENODEV;
248 
249 	if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
250 		/*
251 		 * Use local_paca instead of get_lppaca() since
252 		 * preemption is not disabled, and it is not required in
253 		 * fact, since lppaca_ptr does not need to be the value
254 		 * associated to the current CPU, it can be from any CPU.
255 		 */
256 		if (lppaca_shared_proc(local_paca->lppaca_ptr)) {
257 			cpuidle_state_table = shared_states;
258 			max_idle_state = ARRAY_SIZE(shared_states);
259 		} else {
260 			cpuidle_state_table = dedicated_states;
261 			max_idle_state = ARRAY_SIZE(dedicated_states);
262 		}
263 	} else
264 		return -ENODEV;
265 
266 	if (max_idle_state > 1) {
267 		snooze_timeout_en = true;
268 		snooze_timeout = cpuidle_state_table[1].target_residency *
269 				 tb_ticks_per_usec;
270 	}
271 	return 0;
272 }
273 
274 static int __init pseries_processor_idle_init(void)
275 {
276 	int retval;
277 
278 	retval = pseries_idle_probe();
279 	if (retval)
280 		return retval;
281 
282 	pseries_cpuidle_driver_init();
283 	retval = cpuidle_register(&pseries_idle_driver, NULL);
284 	if (retval) {
285 		printk(KERN_DEBUG "Registration of pseries driver failed.\n");
286 		return retval;
287 	}
288 
289 	retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
290 					   "cpuidle/pseries:online",
291 					   pseries_cpuidle_cpu_online, NULL);
292 	WARN_ON(retval < 0);
293 	retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD,
294 					   "cpuidle/pseries:DEAD", NULL,
295 					   pseries_cpuidle_cpu_dead);
296 	WARN_ON(retval < 0);
297 	printk(KERN_DEBUG "pseries_idle_driver registered\n");
298 	return 0;
299 }
300 
301 device_initcall(pseries_processor_idle_init);
302