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 *) &regs->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}
265