Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Performance event support for s390x
4 *
5 * Copyright IBM Corp. 2012, 2013
6 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
7 */
8#define KMSG_COMPONENT "perf"
9#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
10
11#include <linux/kernel.h>
12#include <linux/perf_event.h>
13#include <linux/kvm_host.h>
14#include <linux/percpu.h>
15#include <linux/export.h>
16#include <linux/seq_file.h>
17#include <linux/spinlock.h>
18#include <linux/uaccess.h>
19#include <linux/compat.h>
20#include <linux/sysfs.h>
21#include <asm/stacktrace.h>
22#include <asm/irq.h>
23#include <asm/cpu_mf.h>
24#include <asm/lowcore.h>
25#include <asm/processor.h>
26#include <asm/sysinfo.h>
27#include <asm/unwind.h>
28
29static struct kvm_s390_sie_block *sie_block(struct pt_regs *regs)
30{
31 struct stack_frame *stack = (struct stack_frame *) regs->gprs[15];
32
33 if (!stack)
34 return NULL;
35
36 return (struct kvm_s390_sie_block *)stack->sie_control_block;
37}
38
39static bool is_in_guest(struct pt_regs *regs)
40{
41 if (user_mode(regs))
42 return false;
43#if IS_ENABLED(CONFIG_KVM)
44 return instruction_pointer(regs) == (unsigned long) &sie_exit;
45#else
46 return false;
47#endif
48}
49
50static unsigned long guest_is_user_mode(struct pt_regs *regs)
51{
52 return sie_block(regs)->gpsw.mask & PSW_MASK_PSTATE;
53}
54
55static unsigned long instruction_pointer_guest(struct pt_regs *regs)
56{
57 return sie_block(regs)->gpsw.addr;
58}
59
60unsigned long perf_instruction_pointer(struct pt_regs *regs)
61{
62 return is_in_guest(regs) ? instruction_pointer_guest(regs)
63 : instruction_pointer(regs);
64}
65
66static unsigned long perf_misc_guest_flags(struct pt_regs *regs)
67{
68 return guest_is_user_mode(regs) ? PERF_RECORD_MISC_GUEST_USER
69 : PERF_RECORD_MISC_GUEST_KERNEL;
70}
71
72static unsigned long perf_misc_flags_sf(struct pt_regs *regs)
73{
74 struct perf_sf_sde_regs *sde_regs;
75 unsigned long flags;
76
77 sde_regs = (struct perf_sf_sde_regs *) ®s->int_parm_long;
78 if (sde_regs->in_guest)
79 flags = user_mode(regs) ? PERF_RECORD_MISC_GUEST_USER
80 : PERF_RECORD_MISC_GUEST_KERNEL;
81 else
82 flags = user_mode(regs) ? PERF_RECORD_MISC_USER
83 : PERF_RECORD_MISC_KERNEL;
84 return flags;
85}
86
87unsigned long perf_misc_flags(struct pt_regs *regs)
88{
89 /* Check if the cpum_sf PMU has created the pt_regs structure.
90 * In this case, perf misc flags can be easily extracted. Otherwise,
91 * do regular checks on the pt_regs content.
92 */
93 if (regs->int_code == 0x1407 && regs->int_parm == CPU_MF_INT_SF_PRA)
94 if (!regs->gprs[15])
95 return perf_misc_flags_sf(regs);
96
97 if (is_in_guest(regs))
98 return perf_misc_guest_flags(regs);
99
100 return user_mode(regs) ? PERF_RECORD_MISC_USER
101 : PERF_RECORD_MISC_KERNEL;
102}
103
104static void print_debug_cf(void)
105{
106 struct cpumf_ctr_info cf_info;
107 int cpu = smp_processor_id();
108
109 memset(&cf_info, 0, sizeof(cf_info));
110 if (!qctri(&cf_info))
111 pr_info("CPU[%i] CPUM_CF: ver=%u.%u A=%04x E=%04x C=%04x\n",
112 cpu, cf_info.cfvn, cf_info.csvn,
113 cf_info.auth_ctl, cf_info.enable_ctl, cf_info.act_ctl);
114}
115
116static void print_debug_sf(void)
117{
118 struct hws_qsi_info_block si;
119 int cpu = smp_processor_id();
120
121 memset(&si, 0, sizeof(si));
122 if (qsi(&si))
123 return;
124
125 pr_info("CPU[%i] CPUM_SF: basic=%i diag=%i min=%lu max=%lu cpu_speed=%u\n",
126 cpu, si.as, si.ad, si.min_sampl_rate, si.max_sampl_rate,
127 si.cpu_speed);
128
129 if (si.as)
130 pr_info("CPU[%i] CPUM_SF: Basic-sampling: a=%i e=%i c=%i"
131 " bsdes=%i tear=%016lx dear=%016lx\n", cpu,
132 si.as, si.es, si.cs, si.bsdes, si.tear, si.dear);
133 if (si.ad)
134 pr_info("CPU[%i] CPUM_SF: Diagnostic-sampling: a=%i e=%i c=%i"
135 " dsdes=%i tear=%016lx dear=%016lx\n", cpu,
136 si.ad, si.ed, si.cd, si.dsdes, si.tear, si.dear);
137}
138
139void perf_event_print_debug(void)
140{
141 unsigned long flags;
142
143 local_irq_save(flags);
144 if (cpum_cf_avail())
145 print_debug_cf();
146 if (cpum_sf_avail())
147 print_debug_sf();
148 local_irq_restore(flags);
149}
150
151/* Service level infrastructure */
152static void sl_print_counter(struct seq_file *m)
153{
154 struct cpumf_ctr_info ci;
155
156 memset(&ci, 0, sizeof(ci));
157 if (qctri(&ci))
158 return;
159
160 seq_printf(m, "CPU-MF: Counter facility: version=%u.%u "
161 "authorization=%04x\n", ci.cfvn, ci.csvn, ci.auth_ctl);
162}
163
164static void sl_print_sampling(struct seq_file *m)
165{
166 struct hws_qsi_info_block si;
167
168 memset(&si, 0, sizeof(si));
169 if (qsi(&si))
170 return;
171
172 if (!si.as && !si.ad)
173 return;
174
175 seq_printf(m, "CPU-MF: Sampling facility: min_rate=%lu max_rate=%lu"
176 " cpu_speed=%u\n", si.min_sampl_rate, si.max_sampl_rate,
177 si.cpu_speed);
178 if (si.as)
179 seq_printf(m, "CPU-MF: Sampling facility: mode=basic"
180 " sample_size=%u\n", si.bsdes);
181 if (si.ad)
182 seq_printf(m, "CPU-MF: Sampling facility: mode=diagnostic"
183 " sample_size=%u\n", si.dsdes);
184}
185
186static void service_level_perf_print(struct seq_file *m,
187 struct service_level *sl)
188{
189 if (cpum_cf_avail())
190 sl_print_counter(m);
191 if (cpum_sf_avail())
192 sl_print_sampling(m);
193}
194
195static struct service_level service_level_perf = {
196 .seq_print = service_level_perf_print,
197};
198
199static int __init service_level_perf_register(void)
200{
201 return register_service_level(&service_level_perf);
202}
203arch_initcall(service_level_perf_register);
204
205void perf_callchain_kernel(struct perf_callchain_entry_ctx *entry,
206 struct pt_regs *regs)
207{
208 struct unwind_state state;
209 unsigned long addr;
210
211 unwind_for_each_frame(&state, current, regs, 0) {
212 addr = unwind_get_return_address(&state);
213 if (!addr || perf_callchain_store(entry, addr))
214 return;
215 }
216}
217
218void perf_callchain_user(struct perf_callchain_entry_ctx *entry,
219 struct pt_regs *regs)
220{
221 struct stack_frame_user __user *sf;
222 unsigned long ip, sp;
223 bool first = true;
224
225 if (is_compat_task())
226 return;
227 perf_callchain_store(entry, instruction_pointer(regs));
228 sf = (void __user *)user_stack_pointer(regs);
229 pagefault_disable();
230 while (entry->nr < entry->max_stack) {
231 if (__get_user(sp, &sf->back_chain))
232 break;
233 if (__get_user(ip, &sf->gprs[8]))
234 break;
235 if (ip & 0x1) {
236 /*
237 * If the instruction address is invalid, and this
238 * is the first stack frame, assume r14 has not
239 * been written to the stack yet. Otherwise exit.
240 */
241 if (first && !(regs->gprs[14] & 0x1))
242 ip = regs->gprs[14];
243 else
244 break;
245 }
246 perf_callchain_store(entry, ip);
247 /* Sanity check: ABI requires SP to be aligned 8 bytes. */
248 if (!sp || sp & 0x7)
249 break;
250 sf = (void __user *)sp;
251 first = false;
252 }
253 pagefault_enable();
254}
255
256/* Perf definitions for PMU event attributes in sysfs */
257ssize_t cpumf_events_sysfs_show(struct device *dev,
258 struct device_attribute *attr, char *page)
259{
260 struct perf_pmu_events_attr *pmu_attr;
261
262 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
263 return sprintf(page, "event=0x%04llx\n", pmu_attr->id);
264}
1/*
2 * Performance event support for s390x
3 *
4 * Copyright IBM Corp. 2012, 2013
5 * Author(s): Hendrik Brueckner <brueckner@linux.vnet.ibm.com>
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License (version 2 only)
9 * as published by the Free Software Foundation.
10 */
11#define KMSG_COMPONENT "perf"
12#define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
13
14#include <linux/kernel.h>
15#include <linux/perf_event.h>
16#include <linux/kvm_host.h>
17#include <linux/percpu.h>
18#include <linux/export.h>
19#include <linux/seq_file.h>
20#include <linux/spinlock.h>
21#include <linux/sysfs.h>
22#include <asm/irq.h>
23#include <asm/cpu_mf.h>
24#include <asm/lowcore.h>
25#include <asm/processor.h>
26#include <asm/sysinfo.h>
27
28const char *perf_pmu_name(void)
29{
30 if (cpum_cf_avail() || cpum_sf_avail())
31 return "CPU-Measurement Facilities (CPU-MF)";
32 return "pmu";
33}
34EXPORT_SYMBOL(perf_pmu_name);
35
36int perf_num_counters(void)
37{
38 int num = 0;
39
40 if (cpum_cf_avail())
41 num += PERF_CPUM_CF_MAX_CTR;
42 if (cpum_sf_avail())
43 num += PERF_CPUM_SF_MAX_CTR;
44
45 return num;
46}
47EXPORT_SYMBOL(perf_num_counters);
48
49static struct kvm_s390_sie_block *sie_block(struct pt_regs *regs)
50{
51 struct stack_frame *stack = (struct stack_frame *) regs->gprs[15];
52
53 if (!stack)
54 return NULL;
55
56 return (struct kvm_s390_sie_block *) stack->empty1[0];
57}
58
59static bool is_in_guest(struct pt_regs *regs)
60{
61 if (user_mode(regs))
62 return false;
63#if IS_ENABLED(CONFIG_KVM)
64 return instruction_pointer(regs) == (unsigned long) &sie_exit;
65#else
66 return false;
67#endif
68}
69
70static unsigned long guest_is_user_mode(struct pt_regs *regs)
71{
72 return sie_block(regs)->gpsw.mask & PSW_MASK_PSTATE;
73}
74
75static unsigned long instruction_pointer_guest(struct pt_regs *regs)
76{
77 return sie_block(regs)->gpsw.addr & PSW_ADDR_INSN;
78}
79
80unsigned long perf_instruction_pointer(struct pt_regs *regs)
81{
82 return is_in_guest(regs) ? instruction_pointer_guest(regs)
83 : instruction_pointer(regs);
84}
85
86static unsigned long perf_misc_guest_flags(struct pt_regs *regs)
87{
88 return guest_is_user_mode(regs) ? PERF_RECORD_MISC_GUEST_USER
89 : PERF_RECORD_MISC_GUEST_KERNEL;
90}
91
92static unsigned long perf_misc_flags_sf(struct pt_regs *regs)
93{
94 struct perf_sf_sde_regs *sde_regs;
95 unsigned long flags;
96
97 sde_regs = (struct perf_sf_sde_regs *) ®s->int_parm_long;
98 if (sde_regs->in_guest)
99 flags = user_mode(regs) ? PERF_RECORD_MISC_GUEST_USER
100 : PERF_RECORD_MISC_GUEST_KERNEL;
101 else
102 flags = user_mode(regs) ? PERF_RECORD_MISC_USER
103 : PERF_RECORD_MISC_KERNEL;
104 return flags;
105}
106
107unsigned long perf_misc_flags(struct pt_regs *regs)
108{
109 /* Check if the cpum_sf PMU has created the pt_regs structure.
110 * In this case, perf misc flags can be easily extracted. Otherwise,
111 * do regular checks on the pt_regs content.
112 */
113 if (regs->int_code == 0x1407 && regs->int_parm == CPU_MF_INT_SF_PRA)
114 if (!regs->gprs[15])
115 return perf_misc_flags_sf(regs);
116
117 if (is_in_guest(regs))
118 return perf_misc_guest_flags(regs);
119
120 return user_mode(regs) ? PERF_RECORD_MISC_USER
121 : PERF_RECORD_MISC_KERNEL;
122}
123
124static void print_debug_cf(void)
125{
126 struct cpumf_ctr_info cf_info;
127 int cpu = smp_processor_id();
128
129 memset(&cf_info, 0, sizeof(cf_info));
130 if (!qctri(&cf_info))
131 pr_info("CPU[%i] CPUM_CF: ver=%u.%u A=%04x E=%04x C=%04x\n",
132 cpu, cf_info.cfvn, cf_info.csvn,
133 cf_info.auth_ctl, cf_info.enable_ctl, cf_info.act_ctl);
134}
135
136static void print_debug_sf(void)
137{
138 struct hws_qsi_info_block si;
139 int cpu = smp_processor_id();
140
141 memset(&si, 0, sizeof(si));
142 if (qsi(&si))
143 return;
144
145 pr_info("CPU[%i] CPUM_SF: basic=%i diag=%i min=%lu max=%lu cpu_speed=%u\n",
146 cpu, si.as, si.ad, si.min_sampl_rate, si.max_sampl_rate,
147 si.cpu_speed);
148
149 if (si.as)
150 pr_info("CPU[%i] CPUM_SF: Basic-sampling: a=%i e=%i c=%i"
151 " bsdes=%i tear=%016lx dear=%016lx\n", cpu,
152 si.as, si.es, si.cs, si.bsdes, si.tear, si.dear);
153 if (si.ad)
154 pr_info("CPU[%i] CPUM_SF: Diagnostic-sampling: a=%i e=%i c=%i"
155 " dsdes=%i tear=%016lx dear=%016lx\n", cpu,
156 si.ad, si.ed, si.cd, si.dsdes, si.tear, si.dear);
157}
158
159void perf_event_print_debug(void)
160{
161 unsigned long flags;
162
163 local_irq_save(flags);
164 if (cpum_cf_avail())
165 print_debug_cf();
166 if (cpum_sf_avail())
167 print_debug_sf();
168 local_irq_restore(flags);
169}
170
171/* Service level infrastructure */
172static void sl_print_counter(struct seq_file *m)
173{
174 struct cpumf_ctr_info ci;
175
176 memset(&ci, 0, sizeof(ci));
177 if (qctri(&ci))
178 return;
179
180 seq_printf(m, "CPU-MF: Counter facility: version=%u.%u "
181 "authorization=%04x\n", ci.cfvn, ci.csvn, ci.auth_ctl);
182}
183
184static void sl_print_sampling(struct seq_file *m)
185{
186 struct hws_qsi_info_block si;
187
188 memset(&si, 0, sizeof(si));
189 if (qsi(&si))
190 return;
191
192 if (!si.as && !si.ad)
193 return;
194
195 seq_printf(m, "CPU-MF: Sampling facility: min_rate=%lu max_rate=%lu"
196 " cpu_speed=%u\n", si.min_sampl_rate, si.max_sampl_rate,
197 si.cpu_speed);
198 if (si.as)
199 seq_printf(m, "CPU-MF: Sampling facility: mode=basic"
200 " sample_size=%u\n", si.bsdes);
201 if (si.ad)
202 seq_printf(m, "CPU-MF: Sampling facility: mode=diagnostic"
203 " sample_size=%u\n", si.dsdes);
204}
205
206static void service_level_perf_print(struct seq_file *m,
207 struct service_level *sl)
208{
209 if (cpum_cf_avail())
210 sl_print_counter(m);
211 if (cpum_sf_avail())
212 sl_print_sampling(m);
213}
214
215static struct service_level service_level_perf = {
216 .seq_print = service_level_perf_print,
217};
218
219static int __init service_level_perf_register(void)
220{
221 return register_service_level(&service_level_perf);
222}
223arch_initcall(service_level_perf_register);
224
225/* See also arch/s390/kernel/traps.c */
226static unsigned long __store_trace(struct perf_callchain_entry *entry,
227 unsigned long sp,
228 unsigned long low, unsigned long high)
229{
230 struct stack_frame *sf;
231 struct pt_regs *regs;
232
233 while (1) {
234 sp = sp & PSW_ADDR_INSN;
235 if (sp < low || sp > high - sizeof(*sf))
236 return sp;
237 sf = (struct stack_frame *) sp;
238 perf_callchain_store(entry, sf->gprs[8] & PSW_ADDR_INSN);
239 /* Follow the backchain. */
240 while (1) {
241 low = sp;
242 sp = sf->back_chain & PSW_ADDR_INSN;
243 if (!sp)
244 break;
245 if (sp <= low || sp > high - sizeof(*sf))
246 return sp;
247 sf = (struct stack_frame *) sp;
248 perf_callchain_store(entry,
249 sf->gprs[8] & PSW_ADDR_INSN);
250 }
251 /* Zero backchain detected, check for interrupt frame. */
252 sp = (unsigned long) (sf + 1);
253 if (sp <= low || sp > high - sizeof(*regs))
254 return sp;
255 regs = (struct pt_regs *) sp;
256 perf_callchain_store(entry, sf->gprs[8] & PSW_ADDR_INSN);
257 low = sp;
258 sp = regs->gprs[15];
259 }
260}
261
262void perf_callchain_kernel(struct perf_callchain_entry *entry,
263 struct pt_regs *regs)
264{
265 unsigned long head;
266 struct stack_frame *head_sf;
267
268 if (user_mode(regs))
269 return;
270
271 head = regs->gprs[15];
272 head_sf = (struct stack_frame *) head;
273
274 if (!head_sf || !head_sf->back_chain)
275 return;
276
277 head = head_sf->back_chain;
278 head = __store_trace(entry, head, S390_lowcore.async_stack - ASYNC_SIZE,
279 S390_lowcore.async_stack);
280
281 __store_trace(entry, head, S390_lowcore.thread_info,
282 S390_lowcore.thread_info + THREAD_SIZE);
283}
284
285/* Perf defintions for PMU event attributes in sysfs */
286ssize_t cpumf_events_sysfs_show(struct device *dev,
287 struct device_attribute *attr, char *page)
288{
289 struct perf_pmu_events_attr *pmu_attr;
290
291 pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
292 return sprintf(page, "event=0x%04llx,name=%s\n",
293 pmu_attr->id, attr->attr.name);
294}
295
296/* Reserve/release functions for sharing perf hardware */
297static DEFINE_SPINLOCK(perf_hw_owner_lock);
298static void *perf_sampling_owner;
299
300int perf_reserve_sampling(void)
301{
302 int err;
303
304 err = 0;
305 spin_lock(&perf_hw_owner_lock);
306 if (perf_sampling_owner) {
307 pr_warn("The sampling facility is already reserved by %p\n",
308 perf_sampling_owner);
309 err = -EBUSY;
310 } else
311 perf_sampling_owner = __builtin_return_address(0);
312 spin_unlock(&perf_hw_owner_lock);
313 return err;
314}
315EXPORT_SYMBOL(perf_reserve_sampling);
316
317void perf_release_sampling(void)
318{
319 spin_lock(&perf_hw_owner_lock);
320 WARN_ON(!perf_sampling_owner);
321 perf_sampling_owner = NULL;
322 spin_unlock(&perf_hw_owner_lock);
323}
324EXPORT_SYMBOL(perf_release_sampling);