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 18static struct lock_contention_bpf *skel; 19 20int 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++] = map__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 bpf_program__set_autoload(skel->progs.collect_lock_syms, false); 155 156 lock_contention_bpf__attach(skel); 157 return 0; 158} 159 160int lock_contention_start(void) 161{ 162 skel->bss->enabled = 1; 163 return 0; 164} 165 166int lock_contention_stop(void) 167{ 168 skel->bss->enabled = 0; 169 return 0; 170} 171 172static const char *lock_contention_get_name(struct lock_contention *con, 173 struct contention_key *key, 174 u64 *stack_trace, u32 flags) 175{ 176 int idx = 0; 177 u64 addr; 178 const char *name = ""; 179 static char name_buf[KSYM_NAME_LEN]; 180 struct symbol *sym; 181 struct map *kmap; 182 struct machine *machine = con->machine; 183 184 if (con->aggr_mode == LOCK_AGGR_TASK) { 185 struct contention_task_data task; 186 int pid = key->pid; 187 int task_fd = bpf_map__fd(skel->maps.task_data); 188 189 /* do not update idle comm which contains CPU number */ 190 if (pid) { 191 struct thread *t = __machine__findnew_thread(machine, /*pid=*/-1, pid); 192 193 if (t == NULL) 194 return name; 195 if (!bpf_map_lookup_elem(task_fd, &pid, &task) && 196 thread__set_comm(t, task.comm, /*timestamp=*/0)) 197 name = task.comm; 198 } 199 return name; 200 } 201 202 if (con->aggr_mode == LOCK_AGGR_ADDR) { 203 int lock_fd = bpf_map__fd(skel->maps.lock_syms); 204 205 /* per-process locks set upper bits of the flags */ 206 if (flags & LCD_F_MMAP_LOCK) 207 return "mmap_lock"; 208 if (flags & LCD_F_SIGHAND_LOCK) 209 return "siglock"; 210 211 /* global locks with symbols */ 212 sym = machine__find_kernel_symbol(machine, key->lock_addr, &kmap); 213 if (sym) 214 return sym->name; 215 216 /* try semi-global locks collected separately */ 217 if (!bpf_map_lookup_elem(lock_fd, &key->lock_addr, &flags)) { 218 if (flags == LOCK_CLASS_RQLOCK) 219 return "rq_lock"; 220 } 221 222 return ""; 223 } 224 225 /* LOCK_AGGR_CALLER: skip lock internal functions */ 226 while (machine__is_lock_function(machine, stack_trace[idx]) && 227 idx < con->max_stack - 1) 228 idx++; 229 230 addr = stack_trace[idx]; 231 sym = machine__find_kernel_symbol(machine, addr, &kmap); 232 233 if (sym) { 234 unsigned long offset; 235 236 offset = map__map_ip(kmap, addr) - sym->start; 237 238 if (offset == 0) 239 return sym->name; 240 241 snprintf(name_buf, sizeof(name_buf), "%s+%#lx", sym->name, offset); 242 } else { 243 snprintf(name_buf, sizeof(name_buf), "%#lx", (unsigned long)addr); 244 } 245 246 return name_buf; 247} 248 249int lock_contention_read(struct lock_contention *con) 250{ 251 int fd, stack, err = 0; 252 struct contention_key *prev_key, key = {}; 253 struct contention_data data = {}; 254 struct lock_stat *st = NULL; 255 struct machine *machine = con->machine; 256 u64 *stack_trace; 257 size_t stack_size = con->max_stack * sizeof(*stack_trace); 258 259 fd = bpf_map__fd(skel->maps.lock_stat); 260 stack = bpf_map__fd(skel->maps.stacks); 261 262 con->fails.task = skel->bss->task_fail; 263 con->fails.stack = skel->bss->stack_fail; 264 con->fails.time = skel->bss->time_fail; 265 con->fails.data = skel->bss->data_fail; 266 267 stack_trace = zalloc(stack_size); 268 if (stack_trace == NULL) 269 return -1; 270 271 if (con->aggr_mode == LOCK_AGGR_TASK) { 272 struct thread *idle = __machine__findnew_thread(machine, 273 /*pid=*/0, 274 /*tid=*/0); 275 thread__set_comm(idle, "swapper", /*timestamp=*/0); 276 } 277 278 if (con->aggr_mode == LOCK_AGGR_ADDR) { 279 DECLARE_LIBBPF_OPTS(bpf_test_run_opts, opts, 280 .flags = BPF_F_TEST_RUN_ON_CPU, 281 ); 282 int prog_fd = bpf_program__fd(skel->progs.collect_lock_syms); 283 284 bpf_prog_test_run_opts(prog_fd, &opts); 285 } 286 287 /* make sure it loads the kernel map */ 288 map__load(maps__first(machine->kmaps)->map); 289 290 prev_key = NULL; 291 while (!bpf_map_get_next_key(fd, prev_key, &key)) { 292 s64 ls_key; 293 const char *name; 294 295 /* to handle errors in the loop body */ 296 err = -1; 297 298 bpf_map_lookup_elem(fd, &key, &data); 299 if (con->save_callstack) { 300 bpf_map_lookup_elem(stack, &key.stack_id, stack_trace); 301 302 if (!match_callstack_filter(machine, stack_trace)) { 303 con->nr_filtered += data.count; 304 goto next; 305 } 306 } 307 308 switch (con->aggr_mode) { 309 case LOCK_AGGR_CALLER: 310 ls_key = key.stack_id; 311 break; 312 case LOCK_AGGR_TASK: 313 ls_key = key.pid; 314 break; 315 case LOCK_AGGR_ADDR: 316 ls_key = key.lock_addr; 317 break; 318 default: 319 goto next; 320 } 321 322 st = lock_stat_find(ls_key); 323 if (st != NULL) { 324 st->wait_time_total += data.total_time; 325 if (st->wait_time_max < data.max_time) 326 st->wait_time_max = data.max_time; 327 if (st->wait_time_min > data.min_time) 328 st->wait_time_min = data.min_time; 329 330 st->nr_contended += data.count; 331 if (st->nr_contended) 332 st->avg_wait_time = st->wait_time_total / st->nr_contended; 333 goto next; 334 } 335 336 name = lock_contention_get_name(con, &key, stack_trace, data.flags); 337 st = lock_stat_findnew(ls_key, name, data.flags); 338 if (st == NULL) 339 break; 340 341 st->nr_contended = data.count; 342 st->wait_time_total = data.total_time; 343 st->wait_time_max = data.max_time; 344 st->wait_time_min = data.min_time; 345 346 if (data.count) 347 st->avg_wait_time = data.total_time / data.count; 348 349 if (con->aggr_mode == LOCK_AGGR_CALLER && verbose > 0) { 350 st->callstack = memdup(stack_trace, stack_size); 351 if (st->callstack == NULL) 352 break; 353 } 354 355next: 356 prev_key = &key; 357 358 /* we're fine now, reset the error */ 359 err = 0; 360 } 361 362 free(stack_trace); 363 364 return err; 365} 366 367int lock_contention_finish(void) 368{ 369 if (skel) { 370 skel->bss->enabled = 0; 371 lock_contention_bpf__destroy(skel); 372 } 373 374 return 0; 375} 376