tools/perf/util/bpf_lock_contention.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "util/debug.h"
   3 #include "util/evlist.h"
   4 #include "util/machine.h"
   5 #include "util/map.h"
   6 #include "util/symbol.h"
   7 #include "util/target.h"
   8 #include "util/thread.h"
   9 #include "util/thread_map.h"
  10 #include "util/lock-contention.h"
  11 #include <linux/zalloc.h>
  12 #include <linux/string.h>
  13 #include <bpf/bpf.h>
  14
  15 #include "bpf_skel/lock_contention.skel.h"
  16 #include "bpf_skel/lock_data.h"
  17
  18 static struct lock_contention_bpf *skel;
  19
  20 int lock_contention_prepare(struct lock_contention *con)
  21 {
  22         int i, fd;
  23         int ncpus = 1, ntasks = 1, ntypes = 1, naddrs = 1;
  24         struct evlist *evlist = con->evlist;
  25         struct target *target = con->target;
  26
  27         skel = lock_contention_bpf__open();
  28         if (!skel) {
  29                 pr_err("Failed to open lock-contention BPF skeleton\n");
  30                 return -1;
  31         }
  32
  33         bpf_map__set_value_size(skel->maps.stacks, con->max_stack * sizeof(u64));
  34         bpf_map__set_max_entries(skel->maps.lock_stat, con->map_nr_entries);
  35         bpf_map__set_max_entries(skel->maps.tstamp, con->map_nr_entries);
  36
  37         if (con->aggr_mode == LOCK_AGGR_TASK)
  38                 bpf_map__set_max_entries(skel->maps.task_data, con->map_nr_entries);
  39         else
  40                 bpf_map__set_max_entries(skel->maps.task_data, 1);
  41
  42         if (con->save_callstack)
  43                 bpf_map__set_max_entries(skel->maps.stacks, con->map_nr_entries);
  44         else
  45                 bpf_map__set_max_entries(skel->maps.stacks, 1);
  46
  47         if (target__has_cpu(target))
  48                 ncpus = perf_cpu_map__nr(evlist->core.user_requested_cpus);
  49         if (target__has_task(target))
  50                 ntasks = perf_thread_map__nr(evlist->core.threads);
  51         if (con->filters->nr_types)
  52                 ntypes = con->filters->nr_types;
  53
  54         /* resolve lock name filters to addr */
  55         if (con->filters->nr_syms) {
  56                 struct symbol *sym;
  57                 struct map *kmap;
  58                 unsigned long *addrs;
  59
  60                 for (i = 0; i < con->filters->nr_syms; i++) {
  61                         sym = machine__find_kernel_symbol_by_name(con->machine,
  62                                                                   con->filters->syms[i],
  63                                                                   &kmap);
  64                         if (sym == NULL) {
  65                                 pr_warning("ignore unknown symbol: %s\n",
  66                                            con->filters->syms[i]);
  67                                 continue;
  68                         }
  69
  70                         addrs = realloc(con->filters->addrs,
  71                                         (con->filters->nr_addrs + 1) * sizeof(*addrs));
  72                         if (addrs == NULL) {
  73                                 pr_warning("memory allocation failure\n");
  74                                 continue;
  75                         }
  76
  77                         addrs[con->filters->nr_addrs++] = kmap->unmap_ip(kmap, sym->start);
  78                         con->filters->addrs = addrs;
  79                 }
  80                 naddrs = con->filters->nr_addrs;
  81         }
  82
  83         bpf_map__set_max_entries(skel->maps.cpu_filter, ncpus);
  84         bpf_map__set_max_entries(skel->maps.task_filter, ntasks);
  85         bpf_map__set_max_entries(skel->maps.type_filter, ntypes);
  86         bpf_map__set_max_entries(skel->maps.addr_filter, naddrs);
  87
  88         if (lock_contention_bpf__load(skel) < 0) {
  89                 pr_err("Failed to load lock-contention BPF skeleton\n");
  90                 return -1;
  91         }
  92
  93         if (target__has_cpu(target)) {
  94                 u32 cpu;
  95                 u8 val = 1;
  96
  97                 skel->bss->has_cpu = 1;
  98                 fd = bpf_map__fd(skel->maps.cpu_filter);
  99
 100                 for (i = 0; i < ncpus; i++) {
 101                         cpu = perf_cpu_map__cpu(evlist->core.user_requested_cpus, i).cpu;
 102                         bpf_map_update_elem(fd, &cpu, &val, BPF_ANY);
 103                 }
 104         }
 105
 106         if (target__has_task(target)) {
 107                 u32 pid;
 108                 u8 val = 1;
 109
 110                 skel->bss->has_task = 1;
 111                 fd = bpf_map__fd(skel->maps.task_filter);
 112
 113                 for (i = 0; i < ntasks; i++) {
 114                         pid = perf_thread_map__pid(evlist->core.threads, i);
 115                         bpf_map_update_elem(fd, &pid, &val, BPF_ANY);
 116                 }
 117         }
 118
 119         if (target__none(target) && evlist->workload.pid > 0) {
 120                 u32 pid = evlist->workload.pid;
 121                 u8 val = 1;
 122
 123                 skel->bss->has_task = 1;
 124                 fd = bpf_map__fd(skel->maps.task_filter);
 125                 bpf_map_update_elem(fd, &pid, &val, BPF_ANY);
 126         }
 127
 128         if (con->filters->nr_types) {
 129                 u8 val = 1;
 130
 131                 skel->bss->has_type = 1;
 132                 fd = bpf_map__fd(skel->maps.type_filter);
 133
 134                 for (i = 0; i < con->filters->nr_types; i++)
 135                         bpf_map_update_elem(fd, &con->filters->types[i], &val, BPF_ANY);
 136         }
 137
 138         if (con->filters->nr_addrs) {
 139                 u8 val = 1;
 140
 141                 skel->bss->has_addr = 1;
 142                 fd = bpf_map__fd(skel->maps.addr_filter);
 143
 144                 for (i = 0; i < con->filters->nr_addrs; i++)
 145                         bpf_map_update_elem(fd, &con->filters->addrs[i], &val, BPF_ANY);
 146         }
 147
 148         /* these don't work well if in the rodata section */
 149         skel->bss->stack_skip = con->stack_skip;
 150         skel->bss->aggr_mode = con->aggr_mode;
 151         skel->bss->needs_callstack = con->save_callstack;
 152         skel->bss->lock_owner = con->owner;
 153
 154         lock_contention_bpf__attach(skel);
 155         return 0;
 156 }
 157
 158 int lock_contention_start(void)
 159 {
 160         skel->bss->enabled = 1;
 161         return 0;
 162 }
 163
 164 int lock_contention_stop(void)
 165 {
 166         skel->bss->enabled = 0;
 167         return 0;
 168 }
 169
 170 static const char *lock_contention_get_name(struct lock_contention *con,
 171                                             struct contention_key *key,
 172                                             u64 *stack_trace, u32 flags)
 173 {
 174         int idx = 0;
 175         u64 addr;
 176         const char *name = "";
 177         static char name_buf[KSYM_NAME_LEN];
 178         struct symbol *sym;
 179         struct map *kmap;
 180         struct machine *machine = con->machine;
 181
 182         if (con->aggr_mode == LOCK_AGGR_TASK) {
 183                 struct contention_task_data task;
 184                 int pid = key->pid;
 185                 int task_fd = bpf_map__fd(skel->maps.task_data);
 186
 187                 /* do not update idle comm which contains CPU number */
 188                 if (pid) {
 189                         struct thread *t = __machine__findnew_thread(machine, /*pid=*/-1, pid);
 190
 191                         if (t == NULL)
 192                                 return name;
 193                         if (!bpf_map_lookup_elem(task_fd, &pid, &task) &&
 194                             thread__set_comm(t, task.comm, /*timestamp=*/0))
 195                                 name = task.comm;
 196                 }
 197                 return name;
 198         }
 199
 200         if (con->aggr_mode == LOCK_AGGR_ADDR) {
 201                 if (flags & LCD_F_MMAP_LOCK)
 202                         return "mmap_lock";
 203                 if (flags & LCD_F_SIGHAND_LOCK)
 204                         return "siglock";
 205                 sym = machine__find_kernel_symbol(machine, key->lock_addr, &kmap);
 206                 if (sym)
 207                         name = sym->name;
 208                 return name;
 209         }
 210
 211         /* LOCK_AGGR_CALLER: skip lock internal functions */
 212         while (machine__is_lock_function(machine, stack_trace[idx]) &&
 213                idx < con->max_stack - 1)
 214                 idx++;
 215
 216         addr = stack_trace[idx];
 217         sym = machine__find_kernel_symbol(machine, addr, &kmap);
 218
 219         if (sym) {
 220                 unsigned long offset;
 221
 222                 offset = kmap->map_ip(kmap, addr) - sym->start;
 223
 224                 if (offset == 0)
 225                         return sym->name;
 226
 227                 snprintf(name_buf, sizeof(name_buf), "%s+%#lx", sym->name, offset);
 228         } else {
 229                 snprintf(name_buf, sizeof(name_buf), "%#lx", (unsigned long)addr);
 230         }
 231
 232         return name_buf;
 233 }
 234
 235 int lock_contention_read(struct lock_contention *con)
 236 {
 237         int fd, stack, err = 0;
 238         struct contention_key *prev_key, key;
 239         struct contention_data data = {};
 240         struct lock_stat *st = NULL;
 241         struct machine *machine = con->machine;
 242         u64 *stack_trace;
 243         size_t stack_size = con->max_stack * sizeof(*stack_trace);
 244
 245         fd = bpf_map__fd(skel->maps.lock_stat);
 246         stack = bpf_map__fd(skel->maps.stacks);
 247
 248         con->lost = skel->bss->lost;
 249
 250         stack_trace = zalloc(stack_size);
 251         if (stack_trace == NULL)
 252                 return -1;
 253
 254         if (con->aggr_mode == LOCK_AGGR_TASK) {
 255                 struct thread *idle = __machine__findnew_thread(machine,
 256                                                                 /*pid=*/0,
 257                                                                 /*tid=*/0);
 258                 thread__set_comm(idle, "swapper", /*timestamp=*/0);
 259         }
 260
 261         /* make sure it loads the kernel map */
 262         map__load(maps__first(machine->kmaps));
 263
 264         prev_key = NULL;
 265         while (!bpf_map_get_next_key(fd, prev_key, &key)) {
 266                 s64 ls_key;
 267                 const char *name;
 268
 269                 /* to handle errors in the loop body */
 270                 err = -1;
 271
 272                 bpf_map_lookup_elem(fd, &key, &data);
 273                 if (con->save_callstack) {
 274                         bpf_map_lookup_elem(stack, &key.stack_id, stack_trace);
 275
 276                         if (!match_callstack_filter(machine, stack_trace))
 277                                 goto next;
 278                 }
 279
 280                 switch (con->aggr_mode) {
 281                 case LOCK_AGGR_CALLER:
 282                         ls_key = key.stack_id;
 283                         break;
 284                 case LOCK_AGGR_TASK:
 285                         ls_key = key.pid;
 286                         break;
 287                 case LOCK_AGGR_ADDR:
 288                         ls_key = key.lock_addr;
 289                         break;
 290                 default:
 291                         goto next;
 292                 }
 293
 294                 st = lock_stat_find(ls_key);
 295                 if (st != NULL) {
 296                         st->wait_time_total += data.total_time;
 297                         if (st->wait_time_max < data.max_time)
 298                                 st->wait_time_max = data.max_time;
 299                         if (st->wait_time_min > data.min_time)
 300                                 st->wait_time_min = data.min_time;
 301
 302                         st->nr_contended += data.count;
 303                         if (st->nr_contended)
 304                                 st->avg_wait_time = st->wait_time_total / st->nr_contended;
 305                         goto next;
 306                 }
 307
 308                 name = lock_contention_get_name(con, &key, stack_trace, data.flags);
 309                 st = lock_stat_findnew(ls_key, name, data.flags);
 310                 if (st == NULL)
 311                         break;
 312
 313                 st->nr_contended = data.count;
 314                 st->wait_time_total = data.total_time;
 315                 st->wait_time_max = data.max_time;
 316                 st->wait_time_min = data.min_time;
 317
 318                 if (data.count)
 319                         st->avg_wait_time = data.total_time / data.count;
 320
 321                 if (con->save_callstack) {
 322                         st->callstack = memdup(stack_trace, stack_size);
 323                         if (st->callstack == NULL)
 324                                 break;
 325                 }
 326
 327 next:
 328                 prev_key = &key;
 329
 330                 /* we're fine now, reset the error */
 331                 err = 0;
 332         }
 333
 334         free(stack_trace);
 335
 336         return err;
 337 }
 338
 339 int lock_contention_finish(void)
 340 {
 341         if (skel) {
 342                 skel->bss->enabled = 0;
 343                 lock_contention_bpf__destroy(skel);
 344         }
 345
 346         return 0;
 347 }