cpuidle-pseries.c

  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/idle.h>
 23#include <asm/plpar_wrappers.h>
 24#include <asm/rtas.h>
 25#include <asm/time.h>
 26
 27static struct cpuidle_driver pseries_idle_driver = {
 28	.name             = "pseries_idle",
 29	.owner            = THIS_MODULE,
 30};
 31
 32static int max_idle_state __read_mostly;
 33static struct cpuidle_state *cpuidle_state_table __read_mostly;
 34static u64 snooze_timeout __read_mostly;
 35static bool snooze_timeout_en __read_mostly;
 36
 37static __cpuidle
 38int snooze_loop(struct cpuidle_device *dev, struct cpuidle_driver *drv,
 39		int index)
 40{
 41	u64 snooze_exit_time;
 42
 43	set_thread_flag(TIF_POLLING_NRFLAG);
 44
 45	pseries_idle_prolog();
 46	raw_local_irq_enable();
 47	snooze_exit_time = get_tb() + snooze_timeout;
 48	dev->poll_time_limit = false;
 49
 50	while (!need_resched()) {
 51		HMT_low();
 52		HMT_very_low();
 53		if (likely(snooze_timeout_en) && get_tb() > snooze_exit_time) {
 54			/*
 55			 * Task has not woken up but we are exiting the polling
 56			 * loop anyway. Require a barrier after polling is
 57			 * cleared to order subsequent test of need_resched().
 58			 */
 59			dev->poll_time_limit = true;
 60			clear_thread_flag(TIF_POLLING_NRFLAG);
 61			smp_mb();
 62			break;
 63		}
 64	}
 65
 66	HMT_medium();
 67	clear_thread_flag(TIF_POLLING_NRFLAG);
 68
 69	raw_local_irq_disable();
 70
 71	pseries_idle_epilog();
 72
 73	return index;
 74}
 75
 76static __cpuidle void check_and_cede_processor(void)
 77{
 78	/*
 79	 * Ensure our interrupt state is properly tracked,
 80	 * also checks if no interrupt has occurred while we
 81	 * were soft-disabled
 82	 */
 83	if (prep_irq_for_idle()) {
 84		cede_processor();
 85#ifdef CONFIG_TRACE_IRQFLAGS
 86		/* Ensure that H_CEDE returns with IRQs on */
 87		if (WARN_ON(!(mfmsr() & MSR_EE)))
 88			__hard_irq_enable();
 89#endif
 90	}
 91}
 92
 93/*
 94 * XCEDE: Extended CEDE states discovered through the
 95 *        "ibm,get-systems-parameter" RTAS call with the token
 96 *        CEDE_LATENCY_TOKEN
 97 */
 98
 99/*
100 * Section 7.3.16 System Parameters Option of PAPR version 2.8.1 has a
101 * table with all the parameters to ibm,get-system-parameters.
102 * CEDE_LATENCY_TOKEN corresponds to the token value for Cede Latency
103 * Settings Information.
104 */
105#define CEDE_LATENCY_TOKEN	45
106
107/*
108 * If the platform supports the cede latency settings information system
109 * parameter it must provide the following information in the NULL terminated
110 * parameter string:
111 *
112 * a. The first byte is the length “N” of each cede latency setting record minus
113 *    one (zero indicates a length of 1 byte).
114 *
115 * b. For each supported cede latency setting a cede latency setting record
116 *    consisting of the first “N” bytes as per the following table.
117 *
118 *    -----------------------------
119 *    | Field           | Field   |
120 *    | Name            | Length  |
121 *    -----------------------------
122 *    | Cede Latency    | 1 Byte  |
123 *    | Specifier Value |         |
124 *    -----------------------------
125 *    | Maximum wakeup  |         |
126 *    | latency in      | 8 Bytes |
127 *    | tb-ticks        |         |
128 *    -----------------------------
129 *    | Responsive to   |         |
130 *    | external        | 1 Byte  |
131 *    | interrupts      |         |
132 *    -----------------------------
133 *
134 * This version has cede latency record size = 10.
135 *
136 * The structure xcede_latency_payload represents a) and b) with
137 * xcede_latency_record representing the table in b).
138 *
139 * xcede_latency_parameter is what gets returned by
140 * ibm,get-systems-parameter RTAS call when made with
141 * CEDE_LATENCY_TOKEN.
142 *
143 * These structures are only used to represent the data obtained by the RTAS
144 * call. The data is in big-endian.
145 */
146struct xcede_latency_record {
147	u8	hint;
148	__be64	latency_ticks;
149	u8	wake_on_irqs;
150} __packed;
151
152// Make space for 16 records, which "should be enough".
153struct xcede_latency_payload {
154	u8     record_size;
155	struct xcede_latency_record records[16];
156} __packed;
157
158struct xcede_latency_parameter {
159	__be16  payload_size;
160	struct xcede_latency_payload payload;
161	u8 null_char;
162} __packed;
163
164static unsigned int nr_xcede_records;
165static struct xcede_latency_parameter xcede_latency_parameter __initdata;
166
167static int __init parse_cede_parameters(void)
168{
169	struct xcede_latency_payload *payload;
170	u32 total_xcede_records_size;
171	u8 xcede_record_size;
172	u16 payload_size;
173	int ret, i;
174
175	ret = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
176			NULL, CEDE_LATENCY_TOKEN, __pa(&xcede_latency_parameter),
177			sizeof(xcede_latency_parameter));
178	if (ret) {
179		pr_err("xcede: Error parsing CEDE_LATENCY_TOKEN\n");
180		return ret;
181	}
182
183	payload_size = be16_to_cpu(xcede_latency_parameter.payload_size);
184	payload = &xcede_latency_parameter.payload;
185
186	xcede_record_size = payload->record_size + 1;
187
188	if (xcede_record_size != sizeof(struct xcede_latency_record)) {
189		pr_err("xcede: Expected record-size %lu. Observed size %u.\n",
190		       sizeof(struct xcede_latency_record), xcede_record_size);
191		return -EINVAL;
192	}
193
194	pr_info("xcede: xcede_record_size = %d\n", xcede_record_size);
195
196	/*
197	 * Since the payload_size includes the last NULL byte and the
198	 * xcede_record_size, the remaining bytes correspond to array of all
199	 * cede_latency settings.
200	 */
201	total_xcede_records_size = payload_size - 2;
202	nr_xcede_records = total_xcede_records_size / xcede_record_size;
203
204	for (i = 0; i < nr_xcede_records; i++) {
205		struct xcede_latency_record *record = &payload->records[i];
206		u64 latency_ticks = be64_to_cpu(record->latency_ticks);
207		u8 wake_on_irqs = record->wake_on_irqs;
208		u8 hint = record->hint;
209
210		pr_info("xcede: Record %d : hint = %u, latency = 0x%llx tb ticks, Wake-on-irq = %u\n",
211			i, hint, latency_ticks, wake_on_irqs);
212	}
213
214	return 0;
215}
216
217#define NR_DEDICATED_STATES	2 /* snooze, CEDE */
218static u8 cede_latency_hint[NR_DEDICATED_STATES];
219
220static __cpuidle
221int dedicated_cede_loop(struct cpuidle_device *dev, struct cpuidle_driver *drv,
222			int index)
223{
224	u8 old_latency_hint;
225
226	pseries_idle_prolog();
227	get_lppaca()->donate_dedicated_cpu = 1;
228	old_latency_hint = get_lppaca()->cede_latency_hint;
229	get_lppaca()->cede_latency_hint = cede_latency_hint[index];
230
231	HMT_medium();
232	check_and_cede_processor();
233
234	raw_local_irq_disable();
235	get_lppaca()->donate_dedicated_cpu = 0;
236	get_lppaca()->cede_latency_hint = old_latency_hint;
237
238	pseries_idle_epilog();
239
240	return index;
241}
242
243static __cpuidle
244int shared_cede_loop(struct cpuidle_device *dev, struct cpuidle_driver *drv,
245		     int index)
246{
247
248	pseries_idle_prolog();
249
250	/*
251	 * Yield the processor to the hypervisor.  We return if
252	 * an external interrupt occurs (which are driven prior
253	 * to returning here) or if a prod occurs from another
254	 * processor. When returning here, external interrupts
255	 * are enabled.
256	 */
257	check_and_cede_processor();
258
259	raw_local_irq_disable();
260	pseries_idle_epilog();
261
262	return index;
263}
264
265/*
266 * States for dedicated partition case.
267 */
268static struct cpuidle_state dedicated_states[NR_DEDICATED_STATES] = {
269	{ /* Snooze */
270		.name = "snooze",
271		.desc = "snooze",
272		.exit_latency = 0,
273		.target_residency = 0,
274		.enter = &snooze_loop,
275		.flags = CPUIDLE_FLAG_POLLING },
276	{ /* CEDE */
277		.name = "CEDE",
278		.desc = "CEDE",
279		.exit_latency = 10,
280		.target_residency = 100,
281		.enter = &dedicated_cede_loop },
282};
283
284/*
285 * States for shared partition case.
286 */
287static struct cpuidle_state shared_states[] = {
288	{ /* Snooze */
289		.name = "snooze",
290		.desc = "snooze",
291		.exit_latency = 0,
292		.target_residency = 0,
293		.enter = &snooze_loop,
294		.flags = CPUIDLE_FLAG_POLLING },
295	{ /* Shared Cede */
296		.name = "Shared Cede",
297		.desc = "Shared Cede",
298		.exit_latency = 10,
299		.target_residency = 100,
300		.enter = &shared_cede_loop },
301};
302
303static int pseries_cpuidle_cpu_online(unsigned int cpu)
304{
305	struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
306
307	if (dev && cpuidle_get_driver()) {
308		cpuidle_pause_and_lock();
309		cpuidle_enable_device(dev);
310		cpuidle_resume_and_unlock();
311	}
312	return 0;
313}
314
315static int pseries_cpuidle_cpu_dead(unsigned int cpu)
316{
317	struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
318
319	if (dev && cpuidle_get_driver()) {
320		cpuidle_pause_and_lock();
321		cpuidle_disable_device(dev);
322		cpuidle_resume_and_unlock();
323	}
324	return 0;
325}
326
327/*
328 * pseries_cpuidle_driver_init()
329 */
330static int pseries_cpuidle_driver_init(void)
331{
332	int idle_state;
333	struct cpuidle_driver *drv = &pseries_idle_driver;
334
335	drv->state_count = 0;
336
337	for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
338		/* Is the state not enabled? */
339		if (cpuidle_state_table[idle_state].enter == NULL)
340			continue;
341
342		drv->states[drv->state_count] =	/* structure copy */
343			cpuidle_state_table[idle_state];
344
345		drv->state_count += 1;
346	}
347
348	return 0;
349}
350
351static void __init fixup_cede0_latency(void)
352{
353	struct xcede_latency_payload *payload;
354	u64 min_xcede_latency_us = UINT_MAX;
355	int i;
356
357	if (parse_cede_parameters())
358		return;
359
360	pr_info("cpuidle: Skipping the %d Extended CEDE idle states\n",
361		nr_xcede_records);
362
363	payload = &xcede_latency_parameter.payload;
364
365	/*
366	 * The CEDE idle state maps to CEDE(0). While the hypervisor
367	 * does not advertise CEDE(0) exit latency values, it does
368	 * advertise the latency values of the extended CEDE states.
369	 * We use the lowest advertised exit latency value as a proxy
370	 * for the exit latency of CEDE(0).
371	 */
372	for (i = 0; i < nr_xcede_records; i++) {
373		struct xcede_latency_record *record = &payload->records[i];
374		u8 hint = record->hint;
375		u64 latency_tb = be64_to_cpu(record->latency_ticks);
376		u64 latency_us = DIV_ROUND_UP_ULL(tb_to_ns(latency_tb), NSEC_PER_USEC);
377
378		/*
379		 * We expect the exit latency of an extended CEDE
380		 * state to be non-zero, it to since it takes at least
381		 * a few nanoseconds to wakeup the idle CPU and
382		 * dispatch the virtual processor into the Linux
383		 * Guest.
384		 *
385		 * So we consider only non-zero value for performing
386		 * the fixup of CEDE(0) latency.
387		 */
388		if (latency_us == 0) {
389			pr_warn("cpuidle: Skipping xcede record %d [hint=%d]. Exit latency = 0us\n",
390				i, hint);
391			continue;
392		}
393
394		if (latency_us < min_xcede_latency_us)
395			min_xcede_latency_us = latency_us;
396	}
397
398	if (min_xcede_latency_us != UINT_MAX) {
399		dedicated_states[1].exit_latency = min_xcede_latency_us;
400		dedicated_states[1].target_residency = 10 * (min_xcede_latency_us);
401		pr_info("cpuidle: Fixed up CEDE exit latency to %llu us\n",
402			min_xcede_latency_us);
403	}
404
405}
406
407/*
408 * pseries_idle_probe()
409 * Choose state table for shared versus dedicated partition
410 */
411static int __init pseries_idle_probe(void)
412{
413
414	if (cpuidle_disable != IDLE_NO_OVERRIDE)
415		return -ENODEV;
416
417	if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
418		if (lppaca_shared_proc()) {
419			cpuidle_state_table = shared_states;
420			max_idle_state = ARRAY_SIZE(shared_states);
421		} else {
422			/*
423			 * Use firmware provided latency values
424			 * starting with POWER10 platforms. In the
425			 * case that we are running on a POWER10
426			 * platform but in an earlier compat mode, we
427			 * can still use the firmware provided values.
428			 *
429			 * However, on platforms prior to POWER10, we
430			 * cannot rely on the accuracy of the firmware
431			 * provided latency values. On such platforms,
432			 * go with the conservative default estimate
433			 * of 10us.
434			 */
435			if (cpu_has_feature(CPU_FTR_ARCH_31) || pvr_version_is(PVR_POWER10))
436				fixup_cede0_latency();
437			cpuidle_state_table = dedicated_states;
438			max_idle_state = NR_DEDICATED_STATES;
439		}
440	} else
441		return -ENODEV;
442
443	if (max_idle_state > 1) {
444		snooze_timeout_en = true;
445		snooze_timeout = cpuidle_state_table[1].target_residency *
446				 tb_ticks_per_usec;
447	}
448	return 0;
449}
450
451static int __init pseries_processor_idle_init(void)
452{
453	int retval;
454
455	retval = pseries_idle_probe();
456	if (retval)
457		return retval;
458
459	pseries_cpuidle_driver_init();
460	retval = cpuidle_register(&pseries_idle_driver, NULL);
461	if (retval) {
462		printk(KERN_DEBUG "Registration of pseries driver failed.\n");
463		return retval;
464	}
465
466	retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
467					   "cpuidle/pseries:online",
468					   pseries_cpuidle_cpu_online, NULL);
469	WARN_ON(retval < 0);
470	retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD,
471					   "cpuidle/pseries:DEAD", NULL,
472					   pseries_cpuidle_cpu_dead);
473	WARN_ON(retval < 0);
474	printk(KERN_DEBUG "pseries_idle_driver registered\n");
475	return 0;
476}
477
478device_initcall(pseries_processor_idle_init);