1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * nd_perf.c: NVDIMM Device Performance Monitoring Unit support
4 *
5 * Perf interface to expose nvdimm performance stats.
6 *
7 * Copyright (C) 2021 IBM Corporation
8 */
9
10#define pr_fmt(fmt) "nvdimm_pmu: " fmt
11
12#include <linux/nd.h>
13#include <linux/platform_device.h>
14
15#define EVENT(_name, _code) enum{_name = _code}
16
17/*
18 * NVDIMM Events codes.
19 */
20
21/* Controller Reset Count */
22EVENT(CTL_RES_CNT, 0x1);
23/* Controller Reset Elapsed Time */
24EVENT(CTL_RES_TM, 0x2);
25/* Power-on Seconds */
26EVENT(POWERON_SECS, 0x3);
27/* Life Remaining */
28EVENT(MEM_LIFE, 0x4);
29/* Critical Resource Utilization */
30EVENT(CRI_RES_UTIL, 0x5);
31/* Host Load Count */
32EVENT(HOST_L_CNT, 0x6);
33/* Host Store Count */
34EVENT(HOST_S_CNT, 0x7);
35/* Host Store Duration */
36EVENT(HOST_S_DUR, 0x8);
37/* Host Load Duration */
38EVENT(HOST_L_DUR, 0x9);
39/* Media Read Count */
40EVENT(MED_R_CNT, 0xa);
41/* Media Write Count */
42EVENT(MED_W_CNT, 0xb);
43/* Media Read Duration */
44EVENT(MED_R_DUR, 0xc);
45/* Media Write Duration */
46EVENT(MED_W_DUR, 0xd);
47/* Cache Read Hit Count */
48EVENT(CACHE_RH_CNT, 0xe);
49/* Cache Write Hit Count */
50EVENT(CACHE_WH_CNT, 0xf);
51/* Fast Write Count */
52EVENT(FAST_W_CNT, 0x10);
53
54NVDIMM_EVENT_ATTR(ctl_res_cnt, CTL_RES_CNT);
55NVDIMM_EVENT_ATTR(ctl_res_tm, CTL_RES_TM);
56NVDIMM_EVENT_ATTR(poweron_secs, POWERON_SECS);
57NVDIMM_EVENT_ATTR(mem_life, MEM_LIFE);
58NVDIMM_EVENT_ATTR(cri_res_util, CRI_RES_UTIL);
59NVDIMM_EVENT_ATTR(host_l_cnt, HOST_L_CNT);
60NVDIMM_EVENT_ATTR(host_s_cnt, HOST_S_CNT);
61NVDIMM_EVENT_ATTR(host_s_dur, HOST_S_DUR);
62NVDIMM_EVENT_ATTR(host_l_dur, HOST_L_DUR);
63NVDIMM_EVENT_ATTR(med_r_cnt, MED_R_CNT);
64NVDIMM_EVENT_ATTR(med_w_cnt, MED_W_CNT);
65NVDIMM_EVENT_ATTR(med_r_dur, MED_R_DUR);
66NVDIMM_EVENT_ATTR(med_w_dur, MED_W_DUR);
67NVDIMM_EVENT_ATTR(cache_rh_cnt, CACHE_RH_CNT);
68NVDIMM_EVENT_ATTR(cache_wh_cnt, CACHE_WH_CNT);
69NVDIMM_EVENT_ATTR(fast_w_cnt, FAST_W_CNT);
70
71static struct attribute *nvdimm_events_attr[] = {
72 NVDIMM_EVENT_PTR(CTL_RES_CNT),
73 NVDIMM_EVENT_PTR(CTL_RES_TM),
74 NVDIMM_EVENT_PTR(POWERON_SECS),
75 NVDIMM_EVENT_PTR(MEM_LIFE),
76 NVDIMM_EVENT_PTR(CRI_RES_UTIL),
77 NVDIMM_EVENT_PTR(HOST_L_CNT),
78 NVDIMM_EVENT_PTR(HOST_S_CNT),
79 NVDIMM_EVENT_PTR(HOST_S_DUR),
80 NVDIMM_EVENT_PTR(HOST_L_DUR),
81 NVDIMM_EVENT_PTR(MED_R_CNT),
82 NVDIMM_EVENT_PTR(MED_W_CNT),
83 NVDIMM_EVENT_PTR(MED_R_DUR),
84 NVDIMM_EVENT_PTR(MED_W_DUR),
85 NVDIMM_EVENT_PTR(CACHE_RH_CNT),
86 NVDIMM_EVENT_PTR(CACHE_WH_CNT),
87 NVDIMM_EVENT_PTR(FAST_W_CNT),
88 NULL
89};
90
91static struct attribute_group nvdimm_pmu_events_group = {
92 .name = "events",
93 .attrs = nvdimm_events_attr,
94};
95
96PMU_FORMAT_ATTR(event, "config:0-4");
97
98static struct attribute *nvdimm_pmu_format_attr[] = {
99 &format_attr_event.attr,
100 NULL,
101};
102
103static struct attribute_group nvdimm_pmu_format_group = {
104 .name = "format",
105 .attrs = nvdimm_pmu_format_attr,
106};
107
108ssize_t nvdimm_events_sysfs_show(struct device *dev,
109 struct device_attribute *attr, char *page)
110{
111 struct perf_pmu_events_attr *pmu_attr;
112
113 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
114
115 return sprintf(buf: page, fmt: "event=0x%02llx\n", pmu_attr->id);
116}
117
118static ssize_t nvdimm_pmu_cpumask_show(struct device *dev,
119 struct device_attribute *attr, char *buf)
120{
121 struct pmu *pmu = dev_get_drvdata(dev);
122 struct nvdimm_pmu *nd_pmu;
123
124 nd_pmu = container_of(pmu, struct nvdimm_pmu, pmu);
125
126 return cpumap_print_to_pagebuf(list: true, buf, cpumask_of(nd_pmu->cpu));
127}
128
129static int nvdimm_pmu_cpu_offline(unsigned int cpu, struct hlist_node *node)
130{
131 struct nvdimm_pmu *nd_pmu;
132 u32 target;
133 int nodeid;
134 const struct cpumask *cpumask;
135
136 nd_pmu = hlist_entry_safe(node, struct nvdimm_pmu, node);
137
138 /* Clear it, incase given cpu is set in nd_pmu->arch_cpumask */
139 cpumask_test_and_clear_cpu(cpu, cpumask: &nd_pmu->arch_cpumask);
140
141 /*
142 * If given cpu is not same as current designated cpu for
143 * counter access, just return.
144 */
145 if (cpu != nd_pmu->cpu)
146 return 0;
147
148 /* Check for any active cpu in nd_pmu->arch_cpumask */
149 target = cpumask_any(&nd_pmu->arch_cpumask);
150
151 /*
152 * Incase we don't have any active cpu in nd_pmu->arch_cpumask,
153 * check in given cpu's numa node list.
154 */
155 if (target >= nr_cpu_ids) {
156 nodeid = cpu_to_node(cpu);
157 cpumask = cpumask_of_node(node: nodeid);
158 target = cpumask_any_but(mask: cpumask, cpu);
159 }
160 nd_pmu->cpu = target;
161
162 /* Migrate nvdimm pmu events to the new target cpu if valid */
163 if (target >= 0 && target < nr_cpu_ids)
164 perf_pmu_migrate_context(pmu: &nd_pmu->pmu, src_cpu: cpu, dst_cpu: target);
165
166 return 0;
167}
168
169static int nvdimm_pmu_cpu_online(unsigned int cpu, struct hlist_node *node)
170{
171 struct nvdimm_pmu *nd_pmu;
172
173 nd_pmu = hlist_entry_safe(node, struct nvdimm_pmu, node);
174
175 if (nd_pmu->cpu >= nr_cpu_ids)
176 nd_pmu->cpu = cpu;
177
178 return 0;
179}
180
181static int create_cpumask_attr_group(struct nvdimm_pmu *nd_pmu)
182{
183 struct perf_pmu_events_attr *pmu_events_attr;
184 struct attribute **attrs_group;
185 struct attribute_group *nvdimm_pmu_cpumask_group;
186
187 pmu_events_attr = kzalloc(size: sizeof(*pmu_events_attr), GFP_KERNEL);
188 if (!pmu_events_attr)
189 return -ENOMEM;
190
191 attrs_group = kzalloc(size: 2 * sizeof(struct attribute *), GFP_KERNEL);
192 if (!attrs_group) {
193 kfree(objp: pmu_events_attr);
194 return -ENOMEM;
195 }
196
197 /* Allocate memory for cpumask attribute group */
198 nvdimm_pmu_cpumask_group = kzalloc(size: sizeof(*nvdimm_pmu_cpumask_group), GFP_KERNEL);
199 if (!nvdimm_pmu_cpumask_group) {
200 kfree(objp: pmu_events_attr);
201 kfree(objp: attrs_group);
202 return -ENOMEM;
203 }
204
205 sysfs_attr_init(&pmu_events_attr->attr.attr);
206 pmu_events_attr->attr.attr.name = "cpumask";
207 pmu_events_attr->attr.attr.mode = 0444;
208 pmu_events_attr->attr.show = nvdimm_pmu_cpumask_show;
209 attrs_group[0] = &pmu_events_attr->attr.attr;
210 attrs_group[1] = NULL;
211
212 nvdimm_pmu_cpumask_group->attrs = attrs_group;
213 nd_pmu->pmu.attr_groups[NVDIMM_PMU_CPUMASK_ATTR] = nvdimm_pmu_cpumask_group;
214 return 0;
215}
216
217static int nvdimm_pmu_cpu_hotplug_init(struct nvdimm_pmu *nd_pmu)
218{
219 int nodeid, rc;
220 const struct cpumask *cpumask;
221
222 /*
223 * Incase of cpu hotplug feature, arch specific code
224 * can provide required cpumask which can be used
225 * to get designatd cpu for counter access.
226 * Check for any active cpu in nd_pmu->arch_cpumask.
227 */
228 if (!cpumask_empty(srcp: &nd_pmu->arch_cpumask)) {
229 nd_pmu->cpu = cpumask_any(&nd_pmu->arch_cpumask);
230 } else {
231 /* pick active cpu from the cpumask of device numa node. */
232 nodeid = dev_to_node(dev: nd_pmu->dev);
233 cpumask = cpumask_of_node(node: nodeid);
234 nd_pmu->cpu = cpumask_any(cpumask);
235 }
236
237 rc = cpuhp_setup_state_multi(state: CPUHP_AP_ONLINE_DYN, name: "perf/nvdimm:online",
238 startup: nvdimm_pmu_cpu_online, teardown: nvdimm_pmu_cpu_offline);
239
240 if (rc < 0)
241 return rc;
242
243 nd_pmu->cpuhp_state = rc;
244
245 /* Register the pmu instance for cpu hotplug */
246 rc = cpuhp_state_add_instance_nocalls(state: nd_pmu->cpuhp_state, node: &nd_pmu->node);
247 if (rc) {
248 cpuhp_remove_multi_state(state: nd_pmu->cpuhp_state);
249 return rc;
250 }
251
252 /* Create cpumask attribute group */
253 rc = create_cpumask_attr_group(nd_pmu);
254 if (rc) {
255 cpuhp_state_remove_instance_nocalls(state: nd_pmu->cpuhp_state, node: &nd_pmu->node);
256 cpuhp_remove_multi_state(state: nd_pmu->cpuhp_state);
257 return rc;
258 }
259
260 return 0;
261}
262
263static void nvdimm_pmu_free_hotplug_memory(struct nvdimm_pmu *nd_pmu)
264{
265 cpuhp_state_remove_instance_nocalls(state: nd_pmu->cpuhp_state, node: &nd_pmu->node);
266 cpuhp_remove_multi_state(state: nd_pmu->cpuhp_state);
267
268 if (nd_pmu->pmu.attr_groups[NVDIMM_PMU_CPUMASK_ATTR])
269 kfree(objp: nd_pmu->pmu.attr_groups[NVDIMM_PMU_CPUMASK_ATTR]->attrs);
270 kfree(objp: nd_pmu->pmu.attr_groups[NVDIMM_PMU_CPUMASK_ATTR]);
271}
272
273int register_nvdimm_pmu(struct nvdimm_pmu *nd_pmu, struct platform_device *pdev)
274{
275 int rc;
276
277 if (!nd_pmu || !pdev)
278 return -EINVAL;
279
280 /* event functions like add/del/read/event_init and pmu name should not be NULL */
281 if (WARN_ON_ONCE(!(nd_pmu->pmu.event_init && nd_pmu->pmu.add &&
282 nd_pmu->pmu.del && nd_pmu->pmu.read && nd_pmu->pmu.name)))
283 return -EINVAL;
284
285 nd_pmu->pmu.attr_groups = kzalloc(size: (NVDIMM_PMU_NULL_ATTR + 1) *
286 sizeof(struct attribute_group *), GFP_KERNEL);
287 if (!nd_pmu->pmu.attr_groups)
288 return -ENOMEM;
289
290 /*
291 * Add platform_device->dev pointer to nvdimm_pmu to access
292 * device data in events functions.
293 */
294 nd_pmu->dev = &pdev->dev;
295
296 /* Fill attribute groups for the nvdimm pmu device */
297 nd_pmu->pmu.attr_groups[NVDIMM_PMU_FORMAT_ATTR] = &nvdimm_pmu_format_group;
298 nd_pmu->pmu.attr_groups[NVDIMM_PMU_EVENT_ATTR] = &nvdimm_pmu_events_group;
299 nd_pmu->pmu.attr_groups[NVDIMM_PMU_NULL_ATTR] = NULL;
300
301 /* Fill attribute group for cpumask */
302 rc = nvdimm_pmu_cpu_hotplug_init(nd_pmu);
303 if (rc) {
304 pr_info("cpu hotplug feature failed for device: %s\n", nd_pmu->pmu.name);
305 kfree(objp: nd_pmu->pmu.attr_groups);
306 return rc;
307 }
308
309 rc = perf_pmu_register(pmu: &nd_pmu->pmu, name: nd_pmu->pmu.name, type: -1);
310 if (rc) {
311 nvdimm_pmu_free_hotplug_memory(nd_pmu);
312 kfree(objp: nd_pmu->pmu.attr_groups);
313 return rc;
314 }
315
316 pr_info("%s NVDIMM performance monitor support registered\n",
317 nd_pmu->pmu.name);
318
319 return 0;
320}
321EXPORT_SYMBOL_GPL(register_nvdimm_pmu);
322
323void unregister_nvdimm_pmu(struct nvdimm_pmu *nd_pmu)
324{
325 perf_pmu_unregister(pmu: &nd_pmu->pmu);
326 nvdimm_pmu_free_hotplug_memory(nd_pmu);
327 kfree(objp: nd_pmu->pmu.attr_groups);
328 kfree(objp: nd_pmu);
329}
330EXPORT_SYMBOL_GPL(unregister_nvdimm_pmu);
331

source code of linux/drivers/nvdimm/nd_perf.c