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
35MODULE_DESCRIPTION("AMD Family 15h CPU processor power monitor");
36MODULE_AUTHOR("Andreas Herrmann <herrmann.der.user@googlemail.com>");
37MODULE_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
62struct 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
81static bool is_carrizo_or_later(void)
82{
83 return boot_cpu_data.x86 == 0x15 && boot_cpu_data.x86_model >= 0x60;
84}
85
86static ssize_t power1_input_show(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}
139static DEVICE_ATTR_RO(power1_input);
140
141static ssize_t power1_crit_show(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}
148static DEVICE_ATTR_RO(power1_crit);
149
150static 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 */
173static 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
215static ssize_t power1_average_show(struct device *dev,
216 struct device_attribute *attr, char *buf)
217{
218 struct fam15h_power_data *data = dev_get_drvdata(dev);
219 u64 prev_cu_acc_power[MAX_CUS], prev_ptsc[MAX_CUS],
220 jdelta[MAX_CUS];
221 u64 tdelta, avg_acc;
222 int cu, cu_num, ret;
223 signed long leftover;
224
225 /*
226 * With the new x86 topology modelling, x86_max_cores is the
227 * compute unit number.
228 */
229 cu_num = boot_cpu_data.x86_max_cores;
230
231 ret = read_registers(data);
232 if (ret)
233 return 0;
234
235 for (cu = 0; cu < cu_num; cu++) {
236 prev_cu_acc_power[cu] = data->cu_acc_power[cu];
237 prev_ptsc[cu] = data->cpu_sw_pwr_ptsc[cu];
238 }
239
240 leftover = schedule_timeout_interruptible(msecs_to_jiffies(data->power_period));
241 if (leftover)
242 return 0;
243
244 ret = read_registers(data);
245 if (ret)
246 return 0;
247
248 for (cu = 0, avg_acc = 0; cu < cu_num; cu++) {
249 /* check if current compute unit is online */
250 if (data->cu_on[cu] == 0)
251 continue;
252
253 if (data->cu_acc_power[cu] < prev_cu_acc_power[cu]) {
254 jdelta[cu] = data->max_cu_acc_power + data->cu_acc_power[cu];
255 jdelta[cu] -= prev_cu_acc_power[cu];
256 } else {
257 jdelta[cu] = data->cu_acc_power[cu] - prev_cu_acc_power[cu];
258 }
259 tdelta = data->cpu_sw_pwr_ptsc[cu] - prev_ptsc[cu];
260 jdelta[cu] *= data->cpu_pwr_sample_ratio * 1000;
261 do_div(jdelta[cu], tdelta);
262
263 /* the unit is microWatt */
264 avg_acc += jdelta[cu];
265 }
266
267 return sprintf(buf, "%llu\n", (unsigned long long)avg_acc);
268}
269static DEVICE_ATTR_RO(power1_average);
270
271static ssize_t power1_average_interval_show(struct device *dev,
272 struct device_attribute *attr,
273 char *buf)
274{
275 struct fam15h_power_data *data = dev_get_drvdata(dev);
276
277 return sprintf(buf, "%lu\n", data->power_period);
278}
279
280static ssize_t power1_average_interval_store(struct device *dev,
281 struct device_attribute *attr,
282 const char *buf, size_t count)
283{
284 struct fam15h_power_data *data = dev_get_drvdata(dev);
285 unsigned long temp;
286 int ret;
287
288 ret = kstrtoul(buf, 10, &temp);
289 if (ret)
290 return ret;
291
292 if (temp > MAX_INTERVAL)
293 return -EINVAL;
294
295 /* the interval value should be greater than 0 */
296 if (temp <= 0)
297 return -EINVAL;
298
299 data->power_period = temp;
300
301 return count;
302}
303static DEVICE_ATTR_RW(power1_average_interval);
304
305static int fam15h_power_init_attrs(struct pci_dev *pdev,
306 struct fam15h_power_data *data)
307{
308 int n = FAM15H_MIN_NUM_ATTRS;
309 struct attribute **fam15h_power_attrs;
310 struct cpuinfo_x86 *c = &boot_cpu_data;
311
312 if (c->x86 == 0x15 &&
313 (c->x86_model <= 0xf ||
314 (c->x86_model >= 0x60 && c->x86_model <= 0x7f)))
315 n += 1;
316
317 /* check if processor supports accumulated power */
318 if (boot_cpu_has(X86_FEATURE_ACC_POWER))
319 n += 2;
320
321 fam15h_power_attrs = devm_kcalloc(&pdev->dev, n,
322 sizeof(*fam15h_power_attrs),
323 GFP_KERNEL);
324
325 if (!fam15h_power_attrs)
326 return -ENOMEM;
327
328 n = 0;
329 fam15h_power_attrs[n++] = &dev_attr_power1_crit.attr;
330 if (c->x86 == 0x15 &&
331 (c->x86_model <= 0xf ||
332 (c->x86_model >= 0x60 && c->x86_model <= 0x7f)))
333 fam15h_power_attrs[n++] = &dev_attr_power1_input.attr;
334
335 if (boot_cpu_has(X86_FEATURE_ACC_POWER)) {
336 fam15h_power_attrs[n++] = &dev_attr_power1_average.attr;
337 fam15h_power_attrs[n++] = &dev_attr_power1_average_interval.attr;
338 }
339
340 data->group.attrs = fam15h_power_attrs;
341
342 return 0;
343}
344
345static bool should_load_on_this_node(struct pci_dev *f4)
346{
347 u32 val;
348
349 pci_bus_read_config_dword(f4->bus, PCI_DEVFN(PCI_SLOT(f4->devfn), 3),
350 REG_NORTHBRIDGE_CAP, &val);
351 if ((val & BIT(29)) && ((val >> 30) & 3))
352 return false;
353
354 return true;
355}
356
357/*
358 * Newer BKDG versions have an updated recommendation on how to properly
359 * initialize the running average range (was: 0xE, now: 0x9). This avoids
360 * counter saturations resulting in bogus power readings.
361 * We correct this value ourselves to cope with older BIOSes.
362 */
363static const struct pci_device_id affected_device[] = {
364 { PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_NB_F4) },
365 { 0 }
366};
367
368static void tweak_runavg_range(struct pci_dev *pdev)
369{
370 u32 val;
371
372 /*
373 * let this quirk apply only to the current version of the
374 * northbridge, since future versions may change the behavior
375 */
376 if (!pci_match_id(affected_device, pdev))
377 return;
378
379 pci_bus_read_config_dword(pdev->bus,
380 PCI_DEVFN(PCI_SLOT(pdev->devfn), 5),
381 REG_TDP_RUNNING_AVERAGE, &val);
382 if ((val & 0xf) != 0xe)
383 return;
384
385 val &= ~0xf;
386 val |= 0x9;
387 pci_bus_write_config_dword(pdev->bus,
388 PCI_DEVFN(PCI_SLOT(pdev->devfn), 5),
389 REG_TDP_RUNNING_AVERAGE, val);
390}
391
392#ifdef CONFIG_PM
393static int fam15h_power_resume(struct pci_dev *pdev)
394{
395 tweak_runavg_range(pdev);
396 return 0;
397}
398#else
399#define fam15h_power_resume NULL
400#endif
401
402static int fam15h_power_init_data(struct pci_dev *f4,
403 struct fam15h_power_data *data)
404{
405 u32 val;
406 u64 tmp;
407 int ret;
408
409 pci_read_config_dword(f4, REG_PROCESSOR_TDP, &val);
410 data->base_tdp = val >> 16;
411 tmp = val & 0xffff;
412
413 pci_bus_read_config_dword(f4->bus, PCI_DEVFN(PCI_SLOT(f4->devfn), 5),
414 REG_TDP_LIMIT3, &val);
415
416 data->tdp_to_watts = ((val & 0x3ff) << 6) | ((val >> 10) & 0x3f);
417 tmp *= data->tdp_to_watts;
418
419 /* result not allowed to be >= 256W */
420 if ((tmp >> 16) >= 256)
421 dev_warn(&f4->dev,
422 "Bogus value for ProcessorPwrWatts (processor_pwr_watts>=%u)\n",
423 (unsigned int) (tmp >> 16));
424
425 /* convert to microWatt */
426 data->processor_pwr_watts = (tmp * 15625) >> 10;
427
428 ret = fam15h_power_init_attrs(f4, data);
429 if (ret)
430 return ret;
431
432
433 /* CPUID Fn8000_0007:EDX[12] indicates to support accumulated power */
434 if (!boot_cpu_has(X86_FEATURE_ACC_POWER))
435 return 0;
436
437 /*
438 * determine the ratio of the compute unit power accumulator
439 * sample period to the PTSC counter period by executing CPUID
440 * Fn8000_0007:ECX
441 */
442 data->cpu_pwr_sample_ratio = cpuid_ecx(0x80000007);
443
444 if (rdmsrl_safe(MSR_F15H_CU_MAX_PWR_ACCUMULATOR, &tmp)) {
445 pr_err("Failed to read max compute unit power accumulator MSR\n");
446 return -ENODEV;
447 }
448
449 data->max_cu_acc_power = tmp;
450
451 /*
452 * Milliseconds are a reasonable interval for the measurement.
453 * But it shouldn't set too long here, because several seconds
454 * would cause the read function to hang. So set default
455 * interval as 10 ms.
456 */
457 data->power_period = 10;
458
459 return read_registers(data);
460}
461
462static int fam15h_power_probe(struct pci_dev *pdev,
463 const struct pci_device_id *id)
464{
465 struct fam15h_power_data *data;
466 struct device *dev = &pdev->dev;
467 struct device *hwmon_dev;
468 int ret;
469
470 /*
471 * though we ignore every other northbridge, we still have to
472 * do the tweaking on _each_ node in MCM processors as the counters
473 * are working hand-in-hand
474 */
475 tweak_runavg_range(pdev);
476
477 if (!should_load_on_this_node(pdev))
478 return -ENODEV;
479
480 data = devm_kzalloc(dev, sizeof(struct fam15h_power_data), GFP_KERNEL);
481 if (!data)
482 return -ENOMEM;
483
484 ret = fam15h_power_init_data(pdev, data);
485 if (ret)
486 return ret;
487
488 data->pdev = pdev;
489
490 data->groups[0] = &data->group;
491
492 hwmon_dev = devm_hwmon_device_register_with_groups(dev, "fam15h_power",
493 data,
494 &data->groups[0]);
495 return PTR_ERR_OR_ZERO(hwmon_dev);
496}
497
498static const struct pci_device_id fam15h_power_id_table[] = {
499 { PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_NB_F4) },
500 { PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_M30H_NB_F4) },
501 { PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_M60H_NB_F4) },
502 { PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_15H_M70H_NB_F4) },
503 { PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_16H_NB_F4) },
504 { PCI_VDEVICE(AMD, PCI_DEVICE_ID_AMD_16H_M30H_NB_F4) },
505 {}
506};
507MODULE_DEVICE_TABLE(pci, fam15h_power_id_table);
508
509static struct pci_driver fam15h_power_driver = {
510 .name = "fam15h_power",
511 .id_table = fam15h_power_id_table,
512 .probe = fam15h_power_probe,
513 .resume = fam15h_power_resume,
514};
515
516module_pci_driver(fam15h_power_driver);
517