xref: /openbmc/linux/drivers/hwmon/fam15h_power.c (revision f3539c12)
1 /*
2  * fam15h_power.c - AMD Family 15h processor power monitoring
3  *
4  * Copyright (c) 2011-2016 Advanced Micro Devices, Inc.
5  * Author: Andreas Herrmann <herrmann.der.user@googlemail.com>
6  *
7  *
8  * This driver is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU General Public License; either
10  * version 2 of the License, or (at your option) any later version.
11  *
12  * This driver is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
15  * See the GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with this driver; if not, see <http://www.gnu.org/licenses/>.
19  */
20 
21 #include <linux/err.h>
22 #include <linux/hwmon.h>
23 #include <linux/hwmon-sysfs.h>
24 #include <linux/init.h>
25 #include <linux/module.h>
26 #include <linux/pci.h>
27 #include <linux/bitops.h>
28 #include <linux/cpu.h>
29 #include <linux/cpumask.h>
30 #include <linux/time.h>
31 #include <linux/sched.h>
32 #include <asm/processor.h>
33 #include <asm/msr.h>
34 
35 MODULE_DESCRIPTION("AMD Family 15h CPU processor power monitor");
36 MODULE_AUTHOR("Andreas Herrmann <herrmann.der.user@googlemail.com>");
37 MODULE_LICENSE("GPL");
38 
39 /* D18F3 */
40 #define REG_NORTHBRIDGE_CAP		0xe8
41 
42 /* D18F4 */
43 #define REG_PROCESSOR_TDP		0x1b8
44 
45 /* D18F5 */
46 #define REG_TDP_RUNNING_AVERAGE		0xe0
47 #define REG_TDP_LIMIT3			0xe8
48 
49 #define FAM15H_MIN_NUM_ATTRS		2
50 #define FAM15H_NUM_GROUPS		2
51 #define MAX_CUS				8
52 
53 /* set maximum interval as 1 second */
54 #define MAX_INTERVAL			1000
55 
56 #define MSR_F15H_CU_PWR_ACCUMULATOR	0xc001007a
57 #define MSR_F15H_CU_MAX_PWR_ACCUMULATOR	0xc001007b
58 #define MSR_F15H_PTSC			0xc0010280
59 
60 #define PCI_DEVICE_ID_AMD_15H_M70H_NB_F4 0x15b4
61 
62 struct fam15h_power_data {
63 	struct pci_dev *pdev;
64 	unsigned int tdp_to_watts;
65 	unsigned int base_tdp;
66 	unsigned int processor_pwr_watts;
67 	unsigned int cpu_pwr_sample_ratio;
68 	const struct attribute_group *groups[FAM15H_NUM_GROUPS];
69 	struct attribute_group group;
70 	/* maximum accumulated power of a compute unit */
71 	u64 max_cu_acc_power;
72 	/* accumulated power of the compute units */
73 	u64 cu_acc_power[MAX_CUS];
74 	/* performance timestamp counter */
75 	u64 cpu_sw_pwr_ptsc[MAX_CUS];
76 	/* online/offline status of current compute unit */
77 	int cu_on[MAX_CUS];
78 	unsigned long power_period;
79 };
80 
81 static bool is_carrizo_or_later(void)
82 {
83 	return boot_cpu_data.x86 == 0x15 && boot_cpu_data.x86_model >= 0x60;
84 }
85 
86 static ssize_t show_power(struct device *dev,
87 			  struct device_attribute *attr, char *buf)
88 {
89 	u32 val, tdp_limit, running_avg_range;
90 	s32 running_avg_capture;
91 	u64 curr_pwr_watts;
92 	struct fam15h_power_data *data = dev_get_drvdata(dev);
93 	struct pci_dev *f4 = data->pdev;
94 
95 	pci_bus_read_config_dword(f4->bus, PCI_DEVFN(PCI_SLOT(f4->devfn), 5),
96 				  REG_TDP_RUNNING_AVERAGE, &val);
97 
98 	/*
99 	 * On Carrizo and later platforms, TdpRunAvgAccCap bit field
100 	 * is extended to 4:31 from 4:25.
101 	 */
102 	if (is_carrizo_or_later()) {
103 		running_avg_capture = val >> 4;
104 		running_avg_capture = sign_extend32(running_avg_capture, 27);
105 	} else {
106 		running_avg_capture = (val >> 4) & 0x3fffff;
107 		running_avg_capture = sign_extend32(running_avg_capture, 21);
108 	}
109 
110 	running_avg_range = (val & 0xf) + 1;
111 
112 	pci_bus_read_config_dword(f4->bus, PCI_DEVFN(PCI_SLOT(f4->devfn), 5),
113 				  REG_TDP_LIMIT3, &val);
114 
115 	/*
116 	 * On Carrizo and later platforms, ApmTdpLimit bit field
117 	 * is extended to 16:31 from 16:28.
118 	 */
119 	if (is_carrizo_or_later())
120 		tdp_limit = val >> 16;
121 	else
122 		tdp_limit = (val >> 16) & 0x1fff;
123 
124 	curr_pwr_watts = ((u64)(tdp_limit +
125 				data->base_tdp)) << running_avg_range;
126 	curr_pwr_watts -= running_avg_capture;
127 	curr_pwr_watts *= data->tdp_to_watts;
128 
129 	/*
130 	 * Convert to microWatt
131 	 *
132 	 * power is in Watt provided as fixed point integer with
133 	 * scaling factor 1/(2^16).  For conversion we use
134 	 * (10^6)/(2^16) = 15625/(2^10)
135 	 */
136 	curr_pwr_watts = (curr_pwr_watts * 15625) >> (10 + running_avg_range);
137 	return sprintf(buf, "%u\n", (unsigned int) curr_pwr_watts);
138 }
139 static DEVICE_ATTR(power1_input, S_IRUGO, show_power, NULL);
140 
141 static ssize_t show_power_crit(struct device *dev,
142 			       struct device_attribute *attr, char *buf)
143 {
144 	struct fam15h_power_data *data = dev_get_drvdata(dev);
145 
146 	return sprintf(buf, "%u\n", data->processor_pwr_watts);
147 }
148 static DEVICE_ATTR(power1_crit, S_IRUGO, show_power_crit, NULL);
149 
150 static void do_read_registers_on_cu(void *_data)
151 {
152 	struct fam15h_power_data *data = _data;
153 	int cpu, cu;
154 
155 	cpu = smp_processor_id();
156 
157 	/*
158 	 * With the new x86 topology modelling, cpu core id actually
159 	 * is compute unit id.
160 	 */
161 	cu = cpu_data(cpu).cpu_core_id;
162 
163 	rdmsrl_safe(MSR_F15H_CU_PWR_ACCUMULATOR, &data->cu_acc_power[cu]);
164 	rdmsrl_safe(MSR_F15H_PTSC, &data->cpu_sw_pwr_ptsc[cu]);
165 
166 	data->cu_on[cu] = 1;
167 }
168 
169 /*
170  * This function is only able to be called when CPUID
171  * Fn8000_0007:EDX[12] is set.
172  */
173 static int read_registers(struct fam15h_power_data *data)
174 {
175 	int core, this_core;
176 	cpumask_var_t mask;
177 	int ret, cpu;
178 
179 	ret = zalloc_cpumask_var(&mask, GFP_KERNEL);
180 	if (!ret)
181 		return -ENOMEM;
182 
183 	memset(data->cu_on, 0, sizeof(int) * MAX_CUS);
184 
185 	get_online_cpus();
186 
187 	/*
188 	 * Choose the first online core of each compute unit, and then
189 	 * read their MSR value of power and ptsc in a single IPI,
190 	 * because the MSR value of CPU core represent the compute
191 	 * unit's.
192 	 */
193 	core = -1;
194 
195 	for_each_online_cpu(cpu) {
196 		this_core = topology_core_id(cpu);
197 
198 		if (this_core == core)
199 			continue;
200 
201 		core = this_core;
202 
203 		/* get any CPU on this compute unit */
204 		cpumask_set_cpu(cpumask_any(topology_sibling_cpumask(cpu)), mask);
205 	}
206 
207 	on_each_cpu_mask(mask, do_read_registers_on_cu, data, true);
208 
209 	put_online_cpus();
210 	free_cpumask_var(mask);
211 
212 	return 0;
213 }
214 
215 static ssize_t acc_show_power(struct device *dev,
216 			      struct device_attribute *attr,
217 			      char *buf)
218 {
219 	struct fam15h_power_data *data = dev_get_drvdata(dev);
220 	u64 prev_cu_acc_power[MAX_CUS], prev_ptsc[MAX_CUS],
221 	    jdelta[MAX_CUS];
222 	u64 tdelta, avg_acc;
223 	int cu, cu_num, ret;
224 	signed long leftover;
225 
226 	/*
227 	 * With the new x86 topology modelling, x86_max_cores is the
228 	 * compute unit number.
229 	 */
230 	cu_num = boot_cpu_data.x86_max_cores;
231 
232 	ret = read_registers(data);
233 	if (ret)
234 		return 0;
235 
236 	for (cu = 0; cu < cu_num; cu++) {
237 		prev_cu_acc_power[cu] = data->cu_acc_power[cu];
238 		prev_ptsc[cu] = data->cpu_sw_pwr_ptsc[cu];
239 	}
240 
241 	leftover = schedule_timeout_interruptible(msecs_to_jiffies(data->power_period));
242 	if (leftover)
243 		return 0;
244 
245 	ret = read_registers(data);
246 	if (ret)
247 		return 0;
248 
249 	for (cu = 0, avg_acc = 0; cu < cu_num; cu++) {
250 		/* check if current compute unit is online */
251 		if (data->cu_on[cu] == 0)
252 			continue;
253 
254 		if (data->cu_acc_power[cu] < prev_cu_acc_power[cu]) {
255 			jdelta[cu] = data->max_cu_acc_power + data->cu_acc_power[cu];
256 			jdelta[cu] -= prev_cu_acc_power[cu];
257 		} else {
258 			jdelta[cu] = data->cu_acc_power[cu] - prev_cu_acc_power[cu];
259 		}
260 		tdelta = data->cpu_sw_pwr_ptsc[cu] - prev_ptsc[cu];
261 		jdelta[cu] *= data->cpu_pwr_sample_ratio * 1000;
262 		do_div(jdelta[cu], tdelta);
263 
264 		/* the unit is microWatt */
265 		avg_acc += jdelta[cu];
266 	}
267 
268 	return sprintf(buf, "%llu\n", (unsigned long long)avg_acc);
269 }
270 static DEVICE_ATTR(power1_average, S_IRUGO, acc_show_power, NULL);
271 
272 static ssize_t acc_show_power_period(struct device *dev,
273 				     struct device_attribute *attr,
274 				     char *buf)
275 {
276 	struct fam15h_power_data *data = dev_get_drvdata(dev);
277 
278 	return sprintf(buf, "%lu\n", data->power_period);
279 }
280 
281 static ssize_t acc_set_power_period(struct device *dev,
282 				    struct device_attribute *attr,
283 				    const char *buf, size_t count)
284 {
285 	struct fam15h_power_data *data = dev_get_drvdata(dev);
286 	unsigned long temp;
287 	int ret;
288 
289 	ret = kstrtoul(buf, 10, &temp);
290 	if (ret)
291 		return ret;
292 
293 	if (temp > MAX_INTERVAL)
294 		return -EINVAL;
295 
296 	/* the interval value should be greater than 0 */
297 	if (temp <= 0)
298 		return -EINVAL;
299 
300 	data->power_period = temp;
301 
302 	return count;
303 }
304 static DEVICE_ATTR(power1_average_interval, S_IRUGO | S_IWUSR,
305 		   acc_show_power_period, acc_set_power_period);
306 
307 static int fam15h_power_init_attrs(struct pci_dev *pdev,
308 				   struct fam15h_power_data *data)
309 {
310 	int n = FAM15H_MIN_NUM_ATTRS;
311 	struct attribute **fam15h_power_attrs;
312 	struct cpuinfo_x86 *c = &boot_cpu_data;
313 
314 	if (c->x86 == 0x15 &&
315 	    (c->x86_model <= 0xf ||
316 	     (c->x86_model >= 0x60 && c->x86_model <= 0x7f)))
317 		n += 1;
318 
319 	/* check if processor supports accumulated power */
320 	if (boot_cpu_has(X86_FEATURE_ACC_POWER))
321 		n += 2;
322 
323 	fam15h_power_attrs = devm_kcalloc(&pdev->dev, n,
324 					  sizeof(*fam15h_power_attrs),
325 					  GFP_KERNEL);
326 
327 	if (!fam15h_power_attrs)
328 		return -ENOMEM;
329 
330 	n = 0;
331 	fam15h_power_attrs[n++] = &dev_attr_power1_crit.attr;
332 	if (c->x86 == 0x15 &&
333 	    (c->x86_model <= 0xf ||
334 	     (c->x86_model >= 0x60 && c->x86_model <= 0x7f)))
335 		fam15h_power_attrs[n++] = &dev_attr_power1_input.attr;
336 
337 	if (boot_cpu_has(X86_FEATURE_ACC_POWER)) {
338 		fam15h_power_attrs[n++] = &dev_attr_power1_average.attr;
339 		fam15h_power_attrs[n++] = &dev_attr_power1_average_interval.attr;
340 	}
341 
342 	data->group.attrs = fam15h_power_attrs;
343 
344 	return 0;
345 }
346 
347 static bool should_load_on_this_node(struct pci_dev *f4)
348 {
349 	u32 val;
350 
351 	pci_bus_read_config_dword(f4->bus, PCI_DEVFN(PCI_SLOT(f4->devfn), 3),
352 				  REG_NORTHBRIDGE_CAP, &val);
353 	if ((val & BIT(29)) && ((val >> 30) & 3))
354 		return false;
355 
356 	return true;
357 }
358 
359 /*
360  * Newer BKDG versions have an updated recommendation on how to properly
361  * initialize the running average range (was: 0xE, now: 0x9). This avoids
362  * counter saturations resulting in bogus power readings.
363  * We correct this value ourselves to cope with older BIOSes.
364  */
365 static const struct pci_device_id affected_device[] = {
366 	{ PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_NB_F4) },
367 	{ 0 }
368 };
369 
370 static void tweak_runavg_range(struct pci_dev *pdev)
371 {
372 	u32 val;
373 
374 	/*
375 	 * let this quirk apply only to the current version of the
376 	 * northbridge, since future versions may change the behavior
377 	 */
378 	if (!pci_match_id(affected_device, pdev))
379 		return;
380 
381 	pci_bus_read_config_dword(pdev->bus,
382 		PCI_DEVFN(PCI_SLOT(pdev->devfn), 5),
383 		REG_TDP_RUNNING_AVERAGE, &val);
384 	if ((val & 0xf) != 0xe)
385 		return;
386 
387 	val &= ~0xf;
388 	val |=  0x9;
389 	pci_bus_write_config_dword(pdev->bus,
390 		PCI_DEVFN(PCI_SLOT(pdev->devfn), 5),
391 		REG_TDP_RUNNING_AVERAGE, val);
392 }
393 
394 #ifdef CONFIG_PM
395 static int fam15h_power_resume(struct pci_dev *pdev)
396 {
397 	tweak_runavg_range(pdev);
398 	return 0;
399 }
400 #else
401 #define fam15h_power_resume NULL
402 #endif
403 
404 static int fam15h_power_init_data(struct pci_dev *f4,
405 				  struct fam15h_power_data *data)
406 {
407 	u32 val;
408 	u64 tmp;
409 	int ret;
410 
411 	pci_read_config_dword(f4, REG_PROCESSOR_TDP, &val);
412 	data->base_tdp = val >> 16;
413 	tmp = val & 0xffff;
414 
415 	pci_bus_read_config_dword(f4->bus, PCI_DEVFN(PCI_SLOT(f4->devfn), 5),
416 				  REG_TDP_LIMIT3, &val);
417 
418 	data->tdp_to_watts = ((val & 0x3ff) << 6) | ((val >> 10) & 0x3f);
419 	tmp *= data->tdp_to_watts;
420 
421 	/* result not allowed to be >= 256W */
422 	if ((tmp >> 16) >= 256)
423 		dev_warn(&f4->dev,
424 			 "Bogus value for ProcessorPwrWatts (processor_pwr_watts>=%u)\n",
425 			 (unsigned int) (tmp >> 16));
426 
427 	/* convert to microWatt */
428 	data->processor_pwr_watts = (tmp * 15625) >> 10;
429 
430 	ret = fam15h_power_init_attrs(f4, data);
431 	if (ret)
432 		return ret;
433 
434 
435 	/* CPUID Fn8000_0007:EDX[12] indicates to support accumulated power */
436 	if (!boot_cpu_has(X86_FEATURE_ACC_POWER))
437 		return 0;
438 
439 	/*
440 	 * determine the ratio of the compute unit power accumulator
441 	 * sample period to the PTSC counter period by executing CPUID
442 	 * Fn8000_0007:ECX
443 	 */
444 	data->cpu_pwr_sample_ratio = cpuid_ecx(0x80000007);
445 
446 	if (rdmsrl_safe(MSR_F15H_CU_MAX_PWR_ACCUMULATOR, &tmp)) {
447 		pr_err("Failed to read max compute unit power accumulator MSR\n");
448 		return -ENODEV;
449 	}
450 
451 	data->max_cu_acc_power = tmp;
452 
453 	/*
454 	 * Milliseconds are a reasonable interval for the measurement.
455 	 * But it shouldn't set too long here, because several seconds
456 	 * would cause the read function to hang. So set default
457 	 * interval as 10 ms.
458 	 */
459 	data->power_period = 10;
460 
461 	return read_registers(data);
462 }
463 
464 static int fam15h_power_probe(struct pci_dev *pdev,
465 			      const struct pci_device_id *id)
466 {
467 	struct fam15h_power_data *data;
468 	struct device *dev = &pdev->dev;
469 	struct device *hwmon_dev;
470 	int ret;
471 
472 	/*
473 	 * though we ignore every other northbridge, we still have to
474 	 * do the tweaking on _each_ node in MCM processors as the counters
475 	 * are working hand-in-hand
476 	 */
477 	tweak_runavg_range(pdev);
478 
479 	if (!should_load_on_this_node(pdev))
480 		return -ENODEV;
481 
482 	data = devm_kzalloc(dev, sizeof(struct fam15h_power_data), GFP_KERNEL);
483 	if (!data)
484 		return -ENOMEM;
485 
486 	ret = fam15h_power_init_data(pdev, data);
487 	if (ret)
488 		return ret;
489 
490 	data->pdev = pdev;
491 
492 	data->groups[0] = &data->group;
493 
494 	hwmon_dev = devm_hwmon_device_register_with_groups(dev, "fam15h_power",
495 							   data,
496 							   &data->groups[0]);
497 	return PTR_ERR_OR_ZERO(hwmon_dev);
498 }
499 
500 static const struct pci_device_id fam15h_power_id_table[] = {
501 	{ PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_NB_F4) },
502 	{ PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_M30H_NB_F4) },
503 	{ PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_M60H_NB_F4) },
504 	{ PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_M70H_NB_F4) },
505 	{ PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_16H_NB_F4) },
506 	{ PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_16H_M30H_NB_F4) },
507 	{}
508 };
509 MODULE_DEVICE_TABLE(pci, fam15h_power_id_table);
510 
511 static struct pci_driver fam15h_power_driver = {
512 	.name = "fam15h_power",
513 	.id_table = fam15h_power_id_table,
514 	.probe = fam15h_power_probe,
515 	.resume = fam15h_power_resume,
516 };
517 
518 module_pci_driver(fam15h_power_driver);
519