forked from ~ljy/RK356X_SDK_RELEASE
blame | history | patch | commit | commitdiff
hc
2023-12-11 072de836f53be56a70cecf70b43ae43b7ce17376 
 kernel/drivers/cpuidle/cpuidle-pseries.c
....@@ -19,9 +19,11 @@
1919 #include <asm/machdep.h>
2020 #include <asm/firmware.h>
2121 #include <asm/runlatch.h>
22
+#include <asm/idle.h>
2223 #include <asm/plpar_wrappers.h>
24
+#include <asm/rtas.h>
2325 
24
-struct cpuidle_driver pseries_idle_driver = {
26
+static struct cpuidle_driver pseries_idle_driver = {
2527 	.name             = "pseries_idle",
2628 	.owner            = THIS_MODULE,
2729 };
....@@ -31,39 +33,15 @@
3133 static u64 snooze_timeout __read_mostly;
3234 static bool snooze_timeout_en __read_mostly;
3335 
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
-
5736 static int snooze_loop(struct cpuidle_device *dev,
5837 			struct cpuidle_driver *drv,
5938 			int index)
6039 {
61
-	unsigned long in_purr;
6240 	u64 snooze_exit_time;
6341 
6442 	set_thread_flag(TIF_POLLING_NRFLAG);
6543 
66
-	idle_loop_prolog(&in_purr);
44
+	pseries_idle_prolog();
6745 	local_irq_enable();
6846 	snooze_exit_time = get_tb() + snooze_timeout;
6947 
....@@ -87,7 +65,7 @@
8765 
8866 	local_irq_disable();
8967 
90
-	idle_loop_epilog(in_purr);
68
+	pseries_idle_epilog();
9169 
9270 	return index;
9371 }
....@@ -109,22 +87,152 @@
10987 	}
11088 }
11189 
90
+/*
91
+ * XCEDE: Extended CEDE states discovered through the
92
+ *        "ibm,get-systems-parameter" RTAS call with the token
93
+ *        CEDE_LATENCY_TOKEN
94
+ */
95
+
96
+/*
97
+ * Section 7.3.16 System Parameters Option of PAPR version 2.8.1 has a
98
+ * table with all the parameters to ibm,get-system-parameters.
99
+ * CEDE_LATENCY_TOKEN corresponds to the token value for Cede Latency
100
+ * Settings Information.
101
+ */
102
+#define CEDE_LATENCY_TOKEN	45
103
+
104
+/*
105
+ * If the platform supports the cede latency settings information system
106
+ * parameter it must provide the following information in the NULL terminated
107
+ * parameter string:
108
+ *
109
+ * a. The first byte is the length “N” of each cede latency setting record minus
110
+ *    one (zero indicates a length of 1 byte).
111
+ *
112
+ * b. For each supported cede latency setting a cede latency setting record
113
+ *    consisting of the first “N” bytes as per the following table.
114
+ *
115
+ *    -----------------------------
116
+ *    | Field           | Field   |
117
+ *    | Name            | Length  |
118
+ *    -----------------------------
119
+ *    | Cede Latency    | 1 Byte  |
120
+ *    | Specifier Value |         |
121
+ *    -----------------------------
122
+ *    | Maximum wakeup  |         |
123
+ *    | latency in      | 8 Bytes |
124
+ *    | tb-ticks        |         |
125
+ *    -----------------------------
126
+ *    | Responsive to   |         |
127
+ *    | external        | 1 Byte  |
128
+ *    | interrupts      |         |
129
+ *    -----------------------------
130
+ *
131
+ * This version has cede latency record size = 10.
132
+ *
133
+ * The structure xcede_latency_payload represents a) and b) with
134
+ * xcede_latency_record representing the table in b).
135
+ *
136
+ * xcede_latency_parameter is what gets returned by
137
+ * ibm,get-systems-parameter RTAS call when made with
138
+ * CEDE_LATENCY_TOKEN.
139
+ *
140
+ * These structures are only used to represent the data obtained by the RTAS
141
+ * call. The data is in big-endian.
142
+ */
143
+struct xcede_latency_record {
144
+	u8	hint;
145
+	__be64	latency_ticks;
146
+	u8	wake_on_irqs;
147
+} __packed;
148
+
149
+// Make space for 16 records, which "should be enough".
150
+struct xcede_latency_payload {
151
+	u8     record_size;
152
+	struct xcede_latency_record records[16];
153
+} __packed;
154
+
155
+struct xcede_latency_parameter {
156
+	__be16  payload_size;
157
+	struct xcede_latency_payload payload;
158
+	u8 null_char;
159
+} __packed;
160
+
161
+static unsigned int nr_xcede_records;
162
+static struct xcede_latency_parameter xcede_latency_parameter __initdata;
163
+
164
+static int __init parse_cede_parameters(void)
165
+{
166
+	struct xcede_latency_payload *payload;
167
+	u32 total_xcede_records_size;
168
+	u8 xcede_record_size;
169
+	u16 payload_size;
170
+	int ret, i;
171
+
172
+	ret = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
173
+			NULL, CEDE_LATENCY_TOKEN, __pa(&xcede_latency_parameter),
174
+			sizeof(xcede_latency_parameter));
175
+	if (ret) {
176
+		pr_err("xcede: Error parsing CEDE_LATENCY_TOKEN\n");
177
+		return ret;
178
+	}
179
+
180
+	payload_size = be16_to_cpu(xcede_latency_parameter.payload_size);
181
+	payload = &xcede_latency_parameter.payload;
182
+
183
+	xcede_record_size = payload->record_size + 1;
184
+
185
+	if (xcede_record_size != sizeof(struct xcede_latency_record)) {
186
+		pr_err("xcede: Expected record-size %lu. Observed size %u.\n",
187
+		       sizeof(struct xcede_latency_record), xcede_record_size);
188
+		return -EINVAL;
189
+	}
190
+
191
+	pr_info("xcede: xcede_record_size = %d\n", xcede_record_size);
192
+
193
+	/*
194
+	 * Since the payload_size includes the last NULL byte and the
195
+	 * xcede_record_size, the remaining bytes correspond to array of all
196
+	 * cede_latency settings.
197
+	 */
198
+	total_xcede_records_size = payload_size - 2;
199
+	nr_xcede_records = total_xcede_records_size / xcede_record_size;
200
+
201
+	for (i = 0; i < nr_xcede_records; i++) {
202
+		struct xcede_latency_record *record = &payload->records[i];
203
+		u64 latency_ticks = be64_to_cpu(record->latency_ticks);
204
+		u8 wake_on_irqs = record->wake_on_irqs;
205
+		u8 hint = record->hint;
206
+
207
+		pr_info("xcede: Record %d : hint = %u, latency = 0x%llx tb ticks, Wake-on-irq = %u\n",
208
+			i, hint, latency_ticks, wake_on_irqs);
209
+	}
210
+
211
+	return 0;
212
+}
213
+
214
+#define NR_DEDICATED_STATES	2 /* snooze, CEDE */
215
+static u8 cede_latency_hint[NR_DEDICATED_STATES];
216
+
112217 static int dedicated_cede_loop(struct cpuidle_device *dev,
113218 				struct cpuidle_driver *drv,
114219 				int index)
115220 {
116
-	unsigned long in_purr;
221
+	u8 old_latency_hint;
117222 
118
-	idle_loop_prolog(&in_purr);
223
+	pseries_idle_prolog();
119224 	get_lppaca()->donate_dedicated_cpu = 1;
225
+	old_latency_hint = get_lppaca()->cede_latency_hint;
226
+	get_lppaca()->cede_latency_hint = cede_latency_hint[index];
120227 
121228 	HMT_medium();
122229 	check_and_cede_processor();
123230 
124231 	local_irq_disable();
125232 	get_lppaca()->donate_dedicated_cpu = 0;
233
+	get_lppaca()->cede_latency_hint = old_latency_hint;
126234 
127
-	idle_loop_epilog(in_purr);
235
+	pseries_idle_epilog();
128236 
129237 	return index;
130238 }
....@@ -133,9 +241,8 @@
133241 			struct cpuidle_driver *drv,
134242 			int index)
135243 {
136
-	unsigned long in_purr;
137244 
138
-	idle_loop_prolog(&in_purr);
245
+	pseries_idle_prolog();
139246 
140247 	/*
141248 	 * Yield the processor to the hypervisor.  We return if
....@@ -147,7 +254,7 @@
147254 	check_and_cede_processor();
148255 
149256 	local_irq_disable();
150
-	idle_loop_epilog(in_purr);
257
+	pseries_idle_epilog();
151258 
152259 	return index;
153260 }
....@@ -155,7 +262,7 @@
155262 /*
156263  * States for dedicated partition case.
157264  */
158
-static struct cpuidle_state dedicated_states[] = {
265
+static struct cpuidle_state dedicated_states[NR_DEDICATED_STATES] = {
159266 	{ /* Snooze */
160267 		.name = "snooze",
161268 		.desc = "snooze",
....@@ -236,11 +343,66 @@
236343 	return 0;
237344 }
238345 
346
+static void __init fixup_cede0_latency(void)
347
+{
348
+	struct xcede_latency_payload *payload;
349
+	u64 min_latency_us;
350
+	int i;
351
+
352
+	min_latency_us = dedicated_states[1].exit_latency; // CEDE latency
353
+
354
+	if (parse_cede_parameters())
355
+		return;
356
+
357
+	pr_info("cpuidle: Skipping the %d Extended CEDE idle states\n",
358
+		nr_xcede_records);
359
+
360
+	payload = &xcede_latency_parameter.payload;
361
+	for (i = 0; i < nr_xcede_records; i++) {
362
+		struct xcede_latency_record *record = &payload->records[i];
363
+		u64 latency_tb = be64_to_cpu(record->latency_ticks);
364
+		u64 latency_us = DIV_ROUND_UP_ULL(tb_to_ns(latency_tb), NSEC_PER_USEC);
365
+
366
+		if (latency_us == 0)
367
+			pr_warn("cpuidle: xcede record %d has an unrealistic latency of 0us.\n", i);
368
+
369
+		if (latency_us < min_latency_us)
370
+			min_latency_us = latency_us;
371
+	}
372
+
373
+	/*
374
+	 * By default, we assume that CEDE(0) has exit latency 10us,
375
+	 * since there is no way for us to query from the platform.
376
+	 *
377
+	 * However, if the wakeup latency of an Extended CEDE state is
378
+	 * smaller than 10us, then we can be sure that CEDE(0)
379
+	 * requires no more than that.
380
+	 *
381
+	 * Perform the fix-up.
382
+	 */
383
+	if (min_latency_us < dedicated_states[1].exit_latency) {
384
+		/*
385
+		 * We set a minimum of 1us wakeup latency for cede0 to
386
+		 * distinguish it from snooze
387
+		 */
388
+		u64 cede0_latency = 1;
389
+
390
+		if (min_latency_us > cede0_latency)
391
+			cede0_latency = min_latency_us - 1;
392
+
393
+		dedicated_states[1].exit_latency = cede0_latency;
394
+		dedicated_states[1].target_residency = 10 * (cede0_latency);
395
+		pr_info("cpuidle: Fixed up CEDE exit latency to %llu us\n",
396
+			cede0_latency);
397
+	}
398
+
399
+}
400
+
239401 /*
240402  * pseries_idle_probe()
241403  * Choose state table for shared versus dedicated partition
242404  */
243
-static int pseries_idle_probe(void)
405
+static int __init pseries_idle_probe(void)
244406 {
245407 
246408 	if (cpuidle_disable != IDLE_NO_OVERRIDE)
....@@ -257,8 +419,23 @@
257419 			cpuidle_state_table = shared_states;
258420 			max_idle_state = ARRAY_SIZE(shared_states);
259421 		} 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();
260437 			cpuidle_state_table = dedicated_states;
261
-			max_idle_state = ARRAY_SIZE(dedicated_states);
438
+			max_idle_state = NR_DEDICATED_STATES;
262439 		}
263440 	} else
264441 		return -ENODEV;