1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Hypervisor supplied "gpci" ("get performance counter info") performance
4 * counter support
5 *
6 * Author: Cody P Schafer <cody@linux.vnet.ibm.com>
7 * Copyright 2014 IBM Corporation.
8 */
9
10 #define pr_fmt(fmt) "hv-gpci: " fmt
11
12 #include <linux/init.h>
13 #include <linux/perf_event.h>
14 #include <asm/firmware.h>
15 #include <asm/hvcall.h>
16 #include <asm/io.h>
17
18 #include "hv-gpci.h"
19 #include "hv-common.h"
20
21 /*
22 * Example usage:
23 * perf stat -e 'hv_gpci/counter_info_version=3,offset=0,length=8,
24 * secondary_index=0,starting_index=0xffffffff,request=0x10/' ...
25 */
26
27 /* u32 */
28 EVENT_DEFINE_RANGE_FORMAT(request, config, 0, 31);
29 /* u32 */
30 /*
31 * Note that starting_index, phys_processor_idx, sibling_part_id,
32 * hw_chip_id, partition_id all refer to the same bit range. They
33 * are basically aliases for the starting_index. The specific alias
34 * used depends on the event. See REQUEST_IDX_KIND in hv-gpci-requests.h
35 */
36 EVENT_DEFINE_RANGE_FORMAT(starting_index, config, 32, 63);
37 EVENT_DEFINE_RANGE_FORMAT_LITE(phys_processor_idx, config, 32, 63);
38 EVENT_DEFINE_RANGE_FORMAT_LITE(sibling_part_id, config, 32, 63);
39 EVENT_DEFINE_RANGE_FORMAT_LITE(hw_chip_id, config, 32, 63);
40 EVENT_DEFINE_RANGE_FORMAT_LITE(partition_id, config, 32, 63);
41
42 /* u16 */
43 EVENT_DEFINE_RANGE_FORMAT(secondary_index, config1, 0, 15);
44 /* u8 */
45 EVENT_DEFINE_RANGE_FORMAT(counter_info_version, config1, 16, 23);
46 /* u8, bytes of data (1-8) */
47 EVENT_DEFINE_RANGE_FORMAT(length, config1, 24, 31);
48 /* u32, byte offset */
49 EVENT_DEFINE_RANGE_FORMAT(offset, config1, 32, 63);
50
51 static cpumask_t hv_gpci_cpumask;
52
53 static struct attribute *format_attrs[] = {
54 &format_attr_request.attr,
55 &format_attr_starting_index.attr,
56 &format_attr_phys_processor_idx.attr,
57 &format_attr_sibling_part_id.attr,
58 &format_attr_hw_chip_id.attr,
59 &format_attr_partition_id.attr,
60 &format_attr_secondary_index.attr,
61 &format_attr_counter_info_version.attr,
62
63 &format_attr_offset.attr,
64 &format_attr_length.attr,
65 NULL,
66 };
67
68 static struct attribute_group format_group = {
69 .name = "format",
70 .attrs = format_attrs,
71 };
72
73 static struct attribute_group event_group = {
74 .name = "events",
75 /* .attrs is set in init */
76 };
77
78 #define HV_CAPS_ATTR(_name, _format) \
79 static ssize_t _name##_show(struct device *dev, \
80 struct device_attribute *attr, \
81 char *page) \
82 { \
83 struct hv_perf_caps caps; \
84 unsigned long hret = hv_perf_caps_get(&caps); \
85 if (hret) \
86 return -EIO; \
87 \
88 return sprintf(page, _format, caps._name); \
89 } \
90 static struct device_attribute hv_caps_attr_##_name = __ATTR_RO(_name)
91
kernel_version_show(struct device * dev,struct device_attribute * attr,char * page)92 static ssize_t kernel_version_show(struct device *dev,
93 struct device_attribute *attr,
94 char *page)
95 {
96 return sprintf(page, "0x%x\n", COUNTER_INFO_VERSION_CURRENT);
97 }
98
cpumask_show(struct device * dev,struct device_attribute * attr,char * buf)99 static ssize_t cpumask_show(struct device *dev,
100 struct device_attribute *attr, char *buf)
101 {
102 return cpumap_print_to_pagebuf(true, buf, &hv_gpci_cpumask);
103 }
104
105 static DEVICE_ATTR_RO(kernel_version);
106 static DEVICE_ATTR_RO(cpumask);
107
108 HV_CAPS_ATTR(version, "0x%x\n");
109 HV_CAPS_ATTR(ga, "%d\n");
110 HV_CAPS_ATTR(expanded, "%d\n");
111 HV_CAPS_ATTR(lab, "%d\n");
112 HV_CAPS_ATTR(collect_privileged, "%d\n");
113
114 static struct attribute *interface_attrs[] = {
115 &dev_attr_kernel_version.attr,
116 &hv_caps_attr_version.attr,
117 &hv_caps_attr_ga.attr,
118 &hv_caps_attr_expanded.attr,
119 &hv_caps_attr_lab.attr,
120 &hv_caps_attr_collect_privileged.attr,
121 NULL,
122 };
123
124 static struct attribute *cpumask_attrs[] = {
125 &dev_attr_cpumask.attr,
126 NULL,
127 };
128
129 static struct attribute_group cpumask_attr_group = {
130 .attrs = cpumask_attrs,
131 };
132
133 static struct attribute_group interface_group = {
134 .name = "interface",
135 .attrs = interface_attrs,
136 };
137
138 static const struct attribute_group *attr_groups[] = {
139 &format_group,
140 &event_group,
141 &interface_group,
142 &cpumask_attr_group,
143 NULL,
144 };
145
146 static DEFINE_PER_CPU(char, hv_gpci_reqb[HGPCI_REQ_BUFFER_SIZE]) __aligned(sizeof(uint64_t));
147
single_gpci_request(u32 req,u32 starting_index,u16 secondary_index,u8 version_in,u32 offset,u8 length,u64 * value)148 static unsigned long single_gpci_request(u32 req, u32 starting_index,
149 u16 secondary_index, u8 version_in, u32 offset, u8 length,
150 u64 *value)
151 {
152 unsigned long ret;
153 size_t i;
154 u64 count;
155 struct hv_gpci_request_buffer *arg;
156
157 arg = (void *)get_cpu_var(hv_gpci_reqb);
158 memset(arg, 0, HGPCI_REQ_BUFFER_SIZE);
159
160 arg->params.counter_request = cpu_to_be32(req);
161 arg->params.starting_index = cpu_to_be32(starting_index);
162 arg->params.secondary_index = cpu_to_be16(secondary_index);
163 arg->params.counter_info_version_in = version_in;
164
165 ret = plpar_hcall_norets(H_GET_PERF_COUNTER_INFO,
166 virt_to_phys(arg), HGPCI_REQ_BUFFER_SIZE);
167 if (ret) {
168 pr_devel("hcall failed: 0x%lx\n", ret);
169 goto out;
170 }
171
172 /*
173 * we verify offset and length are within the zeroed buffer at event
174 * init.
175 */
176 count = 0;
177 for (i = offset; i < offset + length; i++)
178 count |= (u64)(arg->bytes[i]) << ((length - 1 - (i - offset)) * 8);
179
180 *value = count;
181 out:
182 put_cpu_var(hv_gpci_reqb);
183 return ret;
184 }
185
h_gpci_get_value(struct perf_event * event)186 static u64 h_gpci_get_value(struct perf_event *event)
187 {
188 u64 count;
189 unsigned long ret = single_gpci_request(event_get_request(event),
190 event_get_starting_index(event),
191 event_get_secondary_index(event),
192 event_get_counter_info_version(event),
193 event_get_offset(event),
194 event_get_length(event),
195 &count);
196 if (ret)
197 return 0;
198 return count;
199 }
200
h_gpci_event_update(struct perf_event * event)201 static void h_gpci_event_update(struct perf_event *event)
202 {
203 s64 prev;
204 u64 now = h_gpci_get_value(event);
205 prev = local64_xchg(&event->hw.prev_count, now);
206 local64_add(now - prev, &event->count);
207 }
208
h_gpci_event_start(struct perf_event * event,int flags)209 static void h_gpci_event_start(struct perf_event *event, int flags)
210 {
211 local64_set(&event->hw.prev_count, h_gpci_get_value(event));
212 }
213
h_gpci_event_stop(struct perf_event * event,int flags)214 static void h_gpci_event_stop(struct perf_event *event, int flags)
215 {
216 h_gpci_event_update(event);
217 }
218
h_gpci_event_add(struct perf_event * event,int flags)219 static int h_gpci_event_add(struct perf_event *event, int flags)
220 {
221 if (flags & PERF_EF_START)
222 h_gpci_event_start(event, flags);
223
224 return 0;
225 }
226
h_gpci_event_init(struct perf_event * event)227 static int h_gpci_event_init(struct perf_event *event)
228 {
229 u64 count;
230 u8 length;
231
232 /* Not our event */
233 if (event->attr.type != event->pmu->type)
234 return -ENOENT;
235
236 /* config2 is unused */
237 if (event->attr.config2) {
238 pr_devel("config2 set when reserved\n");
239 return -EINVAL;
240 }
241
242 /* no branch sampling */
243 if (has_branch_stack(event))
244 return -EOPNOTSUPP;
245
246 length = event_get_length(event);
247 if (length < 1 || length > 8) {
248 pr_devel("length invalid\n");
249 return -EINVAL;
250 }
251
252 /* last byte within the buffer? */
253 if ((event_get_offset(event) + length) > HGPCI_MAX_DATA_BYTES) {
254 pr_devel("request outside of buffer: %zu > %zu\n",
255 (size_t)event_get_offset(event) + length,
256 HGPCI_MAX_DATA_BYTES);
257 return -EINVAL;
258 }
259
260 /* check if the request works... */
261 if (single_gpci_request(event_get_request(event),
262 event_get_starting_index(event),
263 event_get_secondary_index(event),
264 event_get_counter_info_version(event),
265 event_get_offset(event),
266 length,
267 &count)) {
268 pr_devel("gpci hcall failed\n");
269 return -EINVAL;
270 }
271
272 return 0;
273 }
274
275 static struct pmu h_gpci_pmu = {
276 .task_ctx_nr = perf_invalid_context,
277
278 .name = "hv_gpci",
279 .attr_groups = attr_groups,
280 .event_init = h_gpci_event_init,
281 .add = h_gpci_event_add,
282 .del = h_gpci_event_stop,
283 .start = h_gpci_event_start,
284 .stop = h_gpci_event_stop,
285 .read = h_gpci_event_update,
286 .capabilities = PERF_PMU_CAP_NO_EXCLUDE,
287 };
288
ppc_hv_gpci_cpu_online(unsigned int cpu)289 static int ppc_hv_gpci_cpu_online(unsigned int cpu)
290 {
291 if (cpumask_empty(&hv_gpci_cpumask))
292 cpumask_set_cpu(cpu, &hv_gpci_cpumask);
293
294 return 0;
295 }
296
ppc_hv_gpci_cpu_offline(unsigned int cpu)297 static int ppc_hv_gpci_cpu_offline(unsigned int cpu)
298 {
299 int target;
300
301 /* Check if exiting cpu is used for collecting gpci events */
302 if (!cpumask_test_and_clear_cpu(cpu, &hv_gpci_cpumask))
303 return 0;
304
305 /* Find a new cpu to collect gpci events */
306 target = cpumask_last(cpu_active_mask);
307
308 if (target < 0 || target >= nr_cpu_ids) {
309 pr_err("hv_gpci: CPU hotplug init failed\n");
310 return -1;
311 }
312
313 /* Migrate gpci events to the new target */
314 cpumask_set_cpu(target, &hv_gpci_cpumask);
315 perf_pmu_migrate_context(&h_gpci_pmu, cpu, target);
316
317 return 0;
318 }
319
hv_gpci_cpu_hotplug_init(void)320 static int hv_gpci_cpu_hotplug_init(void)
321 {
322 return cpuhp_setup_state(CPUHP_AP_PERF_POWERPC_HV_GPCI_ONLINE,
323 "perf/powerpc/hv_gcpi:online",
324 ppc_hv_gpci_cpu_online,
325 ppc_hv_gpci_cpu_offline);
326 }
327
hv_gpci_init(void)328 static int hv_gpci_init(void)
329 {
330 int r;
331 unsigned long hret;
332 struct hv_perf_caps caps;
333 struct hv_gpci_request_buffer *arg;
334
335 hv_gpci_assert_offsets_correct();
336
337 if (!firmware_has_feature(FW_FEATURE_LPAR)) {
338 pr_debug("not a virtualized system, not enabling\n");
339 return -ENODEV;
340 }
341
342 hret = hv_perf_caps_get(&caps);
343 if (hret) {
344 pr_debug("could not obtain capabilities, not enabling, rc=%ld\n",
345 hret);
346 return -ENODEV;
347 }
348
349 /* init cpuhotplug */
350 r = hv_gpci_cpu_hotplug_init();
351 if (r)
352 return r;
353
354 /* sampling not supported */
355 h_gpci_pmu.capabilities |= PERF_PMU_CAP_NO_INTERRUPT;
356
357 arg = (void *)get_cpu_var(hv_gpci_reqb);
358 memset(arg, 0, HGPCI_REQ_BUFFER_SIZE);
359
360 /*
361 * hcall H_GET_PERF_COUNTER_INFO populates the output
362 * counter_info_version value based on the system hypervisor.
363 * Pass the counter request 0x10 corresponds to request type
364 * 'Dispatch_timebase_by_processor', to get the supported
365 * counter_info_version.
366 */
367 arg->params.counter_request = cpu_to_be32(0x10);
368
369 r = plpar_hcall_norets(H_GET_PERF_COUNTER_INFO,
370 virt_to_phys(arg), HGPCI_REQ_BUFFER_SIZE);
371 if (r) {
372 pr_devel("hcall failed, can't get supported counter_info_version: 0x%x\n", r);
373 arg->params.counter_info_version_out = 0x8;
374 }
375
376 /*
377 * Use counter_info_version_out value to assign
378 * required hv-gpci event list.
379 */
380 if (arg->params.counter_info_version_out >= 0x8)
381 event_group.attrs = hv_gpci_event_attrs;
382 else
383 event_group.attrs = hv_gpci_event_attrs_v6;
384
385 put_cpu_var(hv_gpci_reqb);
386
387 r = perf_pmu_register(&h_gpci_pmu, h_gpci_pmu.name, -1);
388 if (r)
389 return r;
390
391 return 0;
392 }
393
394 device_initcall(hv_gpci_init);
395