arch/s390/kernel/perf_event.c

   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
  28 const 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 }
  34 EXPORT_SYMBOL(perf_pmu_name);
  35
  36 int 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 }
  47 EXPORT_SYMBOL(perf_num_counters);
  48
  49 static 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
  59 static 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
  70 static unsigned long guest_is_user_mode(struct pt_regs *regs)
  71 {
  72         return sie_block(regs)->gpsw.mask & PSW_MASK_PSTATE;
  73 }
  74
  75 static unsigned long instruction_pointer_guest(struct pt_regs *regs)
  76 {
  77         return sie_block(regs)->gpsw.addr;
  78 }
  79
  80 unsigned 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
  86 static 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
  92 static 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 *) &regs->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
 107 unsigned 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
 124 static 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
 136 static 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
 159 void 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 */
 172 static 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
 184 static 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
 206 static 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
 215 static struct service_level service_level_perf = {
 216         .seq_print = service_level_perf_print,
 217 };
 218
 219 static int __init service_level_perf_register(void)
 220 {
 221         return register_service_level(&service_level_perf);
 222 }
 223 arch_initcall(service_level_perf_register);
 224
 225 /* See also arch/s390/kernel/traps.c */
 226 static 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                 if (sp < low || sp > high - sizeof(*sf))
 235                         return sp;
 236                 sf = (struct stack_frame *) sp;
 237                 perf_callchain_store(entry, sf->gprs[8]);
 238                 /* Follow the backchain. */
 239                 while (1) {
 240                         low = sp;
 241                         sp = sf->back_chain;
 242                         if (!sp)
 243                                 break;
 244                         if (sp <= low || sp > high - sizeof(*sf))
 245                                 return sp;
 246                         sf = (struct stack_frame *) sp;
 247                         perf_callchain_store(entry, sf->gprs[8]);
 248                 }
 249                 /* Zero backchain detected, check for interrupt frame. */
 250                 sp = (unsigned long) (sf + 1);
 251                 if (sp <= low || sp > high - sizeof(*regs))
 252                         return sp;
 253                 regs = (struct pt_regs *) sp;
 254                 perf_callchain_store(entry, sf->gprs[8]);
 255                 low = sp;
 256                 sp = regs->gprs[15];
 257         }
 258 }
 259
 260 void perf_callchain_kernel(struct perf_callchain_entry *entry,
 261                            struct pt_regs *regs)
 262 {
 263         unsigned long head;
 264         struct stack_frame *head_sf;
 265
 266         if (user_mode(regs))
 267                 return;
 268
 269         head = regs->gprs[15];
 270         head_sf = (struct stack_frame *) head;
 271
 272         if (!head_sf || !head_sf->back_chain)
 273                 return;
 274
 275         head = head_sf->back_chain;
 276         head = __store_trace(entry, head, S390_lowcore.async_stack - ASYNC_SIZE,
 277                              S390_lowcore.async_stack);
 278
 279         __store_trace(entry, head, S390_lowcore.thread_info,
 280                       S390_lowcore.thread_info + THREAD_SIZE);
 281 }
 282
 283 /* Perf defintions for PMU event attributes in sysfs */
 284 ssize_t cpumf_events_sysfs_show(struct device *dev,
 285                                 struct device_attribute *attr, char *page)
 286 {
 287         struct perf_pmu_events_attr *pmu_attr;
 288
 289         pmu_attr = container_of(attr, struct perf_pmu_events_attr, attr);
 290         return sprintf(page, "event=0x%04llx,name=%s\n",
 291                        pmu_attr->id, attr->attr.name);
 292 }
 293
 294 /* Reserve/release functions for sharing perf hardware */
 295 static DEFINE_SPINLOCK(perf_hw_owner_lock);
 296 static void *perf_sampling_owner;
 297
 298 int perf_reserve_sampling(void)
 299 {
 300         int err;
 301
 302         err = 0;
 303         spin_lock(&perf_hw_owner_lock);
 304         if (perf_sampling_owner) {
 305                 pr_warn("The sampling facility is already reserved by %p\n",
 306                         perf_sampling_owner);
 307                 err = -EBUSY;
 308         } else
 309                 perf_sampling_owner = __builtin_return_address(0);
 310         spin_unlock(&perf_hw_owner_lock);
 311         return err;
 312 }
 313 EXPORT_SYMBOL(perf_reserve_sampling);
 314
 315 void perf_release_sampling(void)
 316 {
 317         spin_lock(&perf_hw_owner_lock);
 318         WARN_ON(!perf_sampling_owner);
 319         perf_sampling_owner = NULL;
 320         spin_unlock(&perf_hw_owner_lock);
 321 }
 322 EXPORT_SYMBOL(perf_release_sampling);