Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
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
68 /* Avoid double clear when breaking */
69 if (!dev->poll_time_limit)
70 clear_thread_flag(TIF_POLLING_NRFLAG);
71
72 raw_local_irq_disable();
73
74 pseries_idle_epilog();
75
76 return index;
77}
78
79static __cpuidle void check_and_cede_processor(void)
80{
81 /*
82 * Ensure our interrupt state is properly tracked,
83 * also checks if no interrupt has occurred while we
84 * were soft-disabled
85 */
86 if (prep_irq_for_idle()) {
87 cede_processor();
88#ifdef CONFIG_TRACE_IRQFLAGS
89 /* Ensure that H_CEDE returns with IRQs on */
90 if (WARN_ON(!(mfmsr() & MSR_EE)))
91 __hard_irq_enable();
92#endif
93 }
94}
95
96/*
97 * XCEDE: Extended CEDE states discovered through the
98 * "ibm,get-systems-parameter" RTAS call with the token
99 * CEDE_LATENCY_TOKEN
100 */
101
102/*
103 * Section 7.3.16 System Parameters Option of PAPR version 2.8.1 has a
104 * table with all the parameters to ibm,get-system-parameters.
105 * CEDE_LATENCY_TOKEN corresponds to the token value for Cede Latency
106 * Settings Information.
107 */
108#define CEDE_LATENCY_TOKEN 45
109
110/*
111 * If the platform supports the cede latency settings information system
112 * parameter it must provide the following information in the NULL terminated
113 * parameter string:
114 *
115 * a. The first byte is the length “N” of each cede latency setting record minus
116 * one (zero indicates a length of 1 byte).
117 *
118 * b. For each supported cede latency setting a cede latency setting record
119 * consisting of the first “N” bytes as per the following table.
120 *
121 * -----------------------------
122 * | Field | Field |
123 * | Name | Length |
124 * -----------------------------
125 * | Cede Latency | 1 Byte |
126 * | Specifier Value | |
127 * -----------------------------
128 * | Maximum wakeup | |
129 * | latency in | 8 Bytes |
130 * | tb-ticks | |
131 * -----------------------------
132 * | Responsive to | |
133 * | external | 1 Byte |
134 * | interrupts | |
135 * -----------------------------
136 *
137 * This version has cede latency record size = 10.
138 *
139 * The structure xcede_latency_payload represents a) and b) with
140 * xcede_latency_record representing the table in b).
141 *
142 * xcede_latency_parameter is what gets returned by
143 * ibm,get-systems-parameter RTAS call when made with
144 * CEDE_LATENCY_TOKEN.
145 *
146 * These structures are only used to represent the data obtained by the RTAS
147 * call. The data is in big-endian.
148 */
149struct xcede_latency_record {
150 u8 hint;
151 __be64 latency_ticks;
152 u8 wake_on_irqs;
153} __packed;
154
155// Make space for 16 records, which "should be enough".
156struct xcede_latency_payload {
157 u8 record_size;
158 struct xcede_latency_record records[16];
159} __packed;
160
161struct xcede_latency_parameter {
162 __be16 payload_size;
163 struct xcede_latency_payload payload;
164 u8 null_char;
165} __packed;
166
167static unsigned int nr_xcede_records;
168static struct xcede_latency_parameter xcede_latency_parameter __initdata;
169
170static int __init parse_cede_parameters(void)
171{
172 struct xcede_latency_payload *payload;
173 u32 total_xcede_records_size;
174 u8 xcede_record_size;
175 u16 payload_size;
176 int ret, i;
177
178 ret = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
179 NULL, CEDE_LATENCY_TOKEN, __pa(&xcede_latency_parameter),
180 sizeof(xcede_latency_parameter));
181 if (ret) {
182 pr_err("xcede: Error parsing CEDE_LATENCY_TOKEN\n");
183 return ret;
184 }
185
186 payload_size = be16_to_cpu(xcede_latency_parameter.payload_size);
187 payload = &xcede_latency_parameter.payload;
188
189 xcede_record_size = payload->record_size + 1;
190
191 if (xcede_record_size != sizeof(struct xcede_latency_record)) {
192 pr_err("xcede: Expected record-size %lu. Observed size %u.\n",
193 sizeof(struct xcede_latency_record), xcede_record_size);
194 return -EINVAL;
195 }
196
197 pr_info("xcede: xcede_record_size = %d\n", xcede_record_size);
198
199 /*
200 * Since the payload_size includes the last NULL byte and the
201 * xcede_record_size, the remaining bytes correspond to array of all
202 * cede_latency settings.
203 */
204 total_xcede_records_size = payload_size - 2;
205 nr_xcede_records = total_xcede_records_size / xcede_record_size;
206
207 for (i = 0; i < nr_xcede_records; i++) {
208 struct xcede_latency_record *record = &payload->records[i];
209 u64 latency_ticks = be64_to_cpu(record->latency_ticks);
210 u8 wake_on_irqs = record->wake_on_irqs;
211 u8 hint = record->hint;
212
213 pr_info("xcede: Record %d : hint = %u, latency = 0x%llx tb ticks, Wake-on-irq = %u\n",
214 i, hint, latency_ticks, wake_on_irqs);
215 }
216
217 return 0;
218}
219
220#define NR_DEDICATED_STATES 2 /* snooze, CEDE */
221static u8 cede_latency_hint[NR_DEDICATED_STATES];
222
223static __cpuidle
224int dedicated_cede_loop(struct cpuidle_device *dev, struct cpuidle_driver *drv,
225 int index)
226{
227 u8 old_latency_hint;
228
229 pseries_idle_prolog();
230 get_lppaca()->donate_dedicated_cpu = 1;
231 old_latency_hint = get_lppaca()->cede_latency_hint;
232 get_lppaca()->cede_latency_hint = cede_latency_hint[index];
233
234 HMT_medium();
235 check_and_cede_processor();
236
237 raw_local_irq_disable();
238 get_lppaca()->donate_dedicated_cpu = 0;
239 get_lppaca()->cede_latency_hint = old_latency_hint;
240
241 pseries_idle_epilog();
242
243 return index;
244}
245
246static __cpuidle
247int shared_cede_loop(struct cpuidle_device *dev, struct cpuidle_driver *drv,
248 int index)
249{
250
251 pseries_idle_prolog();
252
253 /*
254 * Yield the processor to the hypervisor. We return if
255 * an external interrupt occurs (which are driven prior
256 * to returning here) or if a prod occurs from another
257 * processor. When returning here, external interrupts
258 * are enabled.
259 */
260 check_and_cede_processor();
261
262 raw_local_irq_disable();
263 pseries_idle_epilog();
264
265 return index;
266}
267
268/*
269 * States for dedicated partition case.
270 */
271static struct cpuidle_state dedicated_states[NR_DEDICATED_STATES] = {
272 { /* Snooze */
273 .name = "snooze",
274 .desc = "snooze",
275 .exit_latency = 0,
276 .target_residency = 0,
277 .enter = &snooze_loop,
278 .flags = CPUIDLE_FLAG_POLLING },
279 { /* CEDE */
280 .name = "CEDE",
281 .desc = "CEDE",
282 .exit_latency = 10,
283 .target_residency = 100,
284 .enter = &dedicated_cede_loop },
285};
286
287/*
288 * States for shared partition case.
289 */
290static struct cpuidle_state shared_states[] = {
291 { /* Snooze */
292 .name = "snooze",
293 .desc = "snooze",
294 .exit_latency = 0,
295 .target_residency = 0,
296 .enter = &snooze_loop,
297 .flags = CPUIDLE_FLAG_POLLING },
298 { /* Shared Cede */
299 .name = "Shared Cede",
300 .desc = "Shared Cede",
301 .exit_latency = 10,
302 .target_residency = 100,
303 .enter = &shared_cede_loop },
304};
305
306static int pseries_cpuidle_cpu_online(unsigned int cpu)
307{
308 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
309
310 if (dev && cpuidle_get_driver()) {
311 cpuidle_pause_and_lock();
312 cpuidle_enable_device(dev);
313 cpuidle_resume_and_unlock();
314 }
315 return 0;
316}
317
318static int pseries_cpuidle_cpu_dead(unsigned int cpu)
319{
320 struct cpuidle_device *dev = per_cpu(cpuidle_devices, cpu);
321
322 if (dev && cpuidle_get_driver()) {
323 cpuidle_pause_and_lock();
324 cpuidle_disable_device(dev);
325 cpuidle_resume_and_unlock();
326 }
327 return 0;
328}
329
330/*
331 * pseries_cpuidle_driver_init()
332 */
333static int pseries_cpuidle_driver_init(void)
334{
335 int idle_state;
336 struct cpuidle_driver *drv = &pseries_idle_driver;
337
338 drv->state_count = 0;
339
340 for (idle_state = 0; idle_state < max_idle_state; ++idle_state) {
341 /* Is the state not enabled? */
342 if (cpuidle_state_table[idle_state].enter == NULL)
343 continue;
344
345 drv->states[drv->state_count] = /* structure copy */
346 cpuidle_state_table[idle_state];
347
348 drv->state_count += 1;
349 }
350
351 return 0;
352}
353
354static void __init fixup_cede0_latency(void)
355{
356 struct xcede_latency_payload *payload;
357 u64 min_xcede_latency_us = UINT_MAX;
358 int i;
359
360 if (parse_cede_parameters())
361 return;
362
363 pr_info("cpuidle: Skipping the %d Extended CEDE idle states\n",
364 nr_xcede_records);
365
366 payload = &xcede_latency_parameter.payload;
367
368 /*
369 * The CEDE idle state maps to CEDE(0). While the hypervisor
370 * does not advertise CEDE(0) exit latency values, it does
371 * advertise the latency values of the extended CEDE states.
372 * We use the lowest advertised exit latency value as a proxy
373 * for the exit latency of CEDE(0).
374 */
375 for (i = 0; i < nr_xcede_records; i++) {
376 struct xcede_latency_record *record = &payload->records[i];
377 u8 hint = record->hint;
378 u64 latency_tb = be64_to_cpu(record->latency_ticks);
379 u64 latency_us = DIV_ROUND_UP_ULL(tb_to_ns(latency_tb), NSEC_PER_USEC);
380
381 /*
382 * We expect the exit latency of an extended CEDE
383 * state to be non-zero, it to since it takes at least
384 * a few nanoseconds to wakeup the idle CPU and
385 * dispatch the virtual processor into the Linux
386 * Guest.
387 *
388 * So we consider only non-zero value for performing
389 * the fixup of CEDE(0) latency.
390 */
391 if (latency_us == 0) {
392 pr_warn("cpuidle: Skipping xcede record %d [hint=%d]. Exit latency = 0us\n",
393 i, hint);
394 continue;
395 }
396
397 if (latency_us < min_xcede_latency_us)
398 min_xcede_latency_us = latency_us;
399 }
400
401 if (min_xcede_latency_us != UINT_MAX) {
402 dedicated_states[1].exit_latency = min_xcede_latency_us;
403 dedicated_states[1].target_residency = 10 * (min_xcede_latency_us);
404 pr_info("cpuidle: Fixed up CEDE exit latency to %llu us\n",
405 min_xcede_latency_us);
406 }
407
408}
409
410/*
411 * pseries_idle_probe()
412 * Choose state table for shared versus dedicated partition
413 */
414static int __init pseries_idle_probe(void)
415{
416
417 if (cpuidle_disable != IDLE_NO_OVERRIDE)
418 return -ENODEV;
419
420 if (firmware_has_feature(FW_FEATURE_SPLPAR)) {
421 if (lppaca_shared_proc()) {
422 cpuidle_state_table = shared_states;
423 max_idle_state = ARRAY_SIZE(shared_states);
424 } else {
425 /*
426 * Use firmware provided latency values
427 * starting with POWER10 platforms. In the
428 * case that we are running on a POWER10
429 * platform but in an earlier compat mode, we
430 * can still use the firmware provided values.
431 *
432 * However, on platforms prior to POWER10, we
433 * cannot rely on the accuracy of the firmware
434 * provided latency values. On such platforms,
435 * go with the conservative default estimate
436 * of 10us.
437 */
438 if (cpu_has_feature(CPU_FTR_ARCH_31) || pvr_version_is(PVR_POWER10))
439 fixup_cede0_latency();
440 cpuidle_state_table = dedicated_states;
441 max_idle_state = NR_DEDICATED_STATES;
442 }
443 } else
444 return -ENODEV;
445
446 if (max_idle_state > 1) {
447 snooze_timeout_en = true;
448 snooze_timeout = cpuidle_state_table[1].target_residency *
449 tb_ticks_per_usec;
450 }
451 return 0;
452}
453
454static int __init pseries_processor_idle_init(void)
455{
456 int retval;
457
458 retval = pseries_idle_probe();
459 if (retval)
460 return retval;
461
462 pseries_cpuidle_driver_init();
463 retval = cpuidle_register(&pseries_idle_driver, NULL);
464 if (retval) {
465 printk(KERN_DEBUG "Registration of pseries driver failed.\n");
466 return retval;
467 }
468
469 retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
470 "cpuidle/pseries:online",
471 pseries_cpuidle_cpu_online, NULL);
472 WARN_ON(retval < 0);
473 retval = cpuhp_setup_state_nocalls(CPUHP_CPUIDLE_DEAD,
474 "cpuidle/pseries:DEAD", NULL,
475 pseries_cpuidle_cpu_dead);
476 WARN_ON(retval < 0);
477 printk(KERN_DEBUG "pseries_idle_driver registered\n");
478 return 0;
479}
480
481device_initcall(pseries_processor_idle_init);