| 1 | /* |
| 2 | * This program is free software; you can redistribute it and/or modify |
| 3 | * it under the terms of the GNU General Public License as published by |
| 4 | * the Free Software Foundation; either version 2 of the License, or |
| 5 | * (at your option) any later version. |
| 6 | * |
| 7 | * This program is distributed in the hope that it will be useful, |
| 8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 10 | * GNU General Public License for more details. |
| 11 | * |
| 12 | * You should have received a copy of the GNU General Public License |
| 13 | * along with this program; if not, write to the Free Software |
| 14 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. |
| 15 | * |
| 16 | * Copyright (C) 2007 Alan Stern |
| 17 | * Copyright (C) IBM Corporation, 2009 |
| 18 | * Copyright (C) 2009, Frederic Weisbecker <fweisbec@gmail.com> |
| 19 | * |
| 20 | * Thanks to Ingo Molnar for his many suggestions. |
| 21 | * |
| 22 | * Authors: Alan Stern <stern@rowland.harvard.edu> |
| 23 | * K.Prasad <prasad@linux.vnet.ibm.com> |
| 24 | * Frederic Weisbecker <fweisbec@gmail.com> |
| 25 | */ |
| 26 | |
| 27 | /* |
| 28 | * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility, |
| 29 | * using the CPU's debug registers. |
| 30 | * This file contains the arch-independent routines. |
| 31 | */ |
| 32 | |
| 33 | #include <linux/irqflags.h> |
| 34 | #include <linux/kallsyms.h> |
| 35 | #include <linux/notifier.h> |
| 36 | #include <linux/kprobes.h> |
| 37 | #include <linux/kdebug.h> |
| 38 | #include <linux/kernel.h> |
| 39 | #include <linux/module.h> |
| 40 | #include <linux/percpu.h> |
| 41 | #include <linux/sched.h> |
| 42 | #include <linux/init.h> |
| 43 | #include <linux/cpu.h> |
| 44 | #include <linux/smp.h> |
| 45 | |
| 46 | #include <linux/hw_breakpoint.h> |
| 47 | |
| 48 | /* |
| 49 | * Constraints data |
| 50 | */ |
| 51 | |
| 52 | /* Number of pinned cpu breakpoints in a cpu */ |
| 53 | static DEFINE_PER_CPU(unsigned int, nr_cpu_bp_pinned); |
| 54 | |
| 55 | /* Number of pinned task breakpoints in a cpu */ |
| 56 | static DEFINE_PER_CPU(unsigned int, nr_task_bp_pinned[HBP_NUM]); |
| 57 | |
| 58 | /* Number of non-pinned cpu/task breakpoints in a cpu */ |
| 59 | static DEFINE_PER_CPU(unsigned int, nr_bp_flexible); |
| 60 | |
| 61 | /* Gather the number of total pinned and un-pinned bp in a cpuset */ |
| 62 | struct bp_busy_slots { |
| 63 | unsigned int pinned; |
| 64 | unsigned int flexible; |
| 65 | }; |
| 66 | |
| 67 | /* Serialize accesses to the above constraints */ |
| 68 | static DEFINE_MUTEX(nr_bp_mutex); |
| 69 | |
| 70 | /* |
| 71 | * Report the maximum number of pinned breakpoints a task |
| 72 | * have in this cpu |
| 73 | */ |
| 74 | static unsigned int max_task_bp_pinned(int cpu) |
| 75 | { |
| 76 | int i; |
| 77 | unsigned int *tsk_pinned = per_cpu(nr_task_bp_pinned, cpu); |
| 78 | |
| 79 | for (i = HBP_NUM -1; i >= 0; i--) { |
| 80 | if (tsk_pinned[i] > 0) |
| 81 | return i + 1; |
| 82 | } |
| 83 | |
| 84 | return 0; |
| 85 | } |
| 86 | |
| 87 | static int task_bp_pinned(struct task_struct *tsk) |
| 88 | { |
| 89 | struct perf_event_context *ctx = tsk->perf_event_ctxp; |
| 90 | struct list_head *list; |
| 91 | struct perf_event *bp; |
| 92 | unsigned long flags; |
| 93 | int count = 0; |
| 94 | |
| 95 | if (WARN_ONCE(!ctx, "No perf context for this task")) |
| 96 | return 0; |
| 97 | |
| 98 | list = &ctx->event_list; |
| 99 | |
| 100 | raw_spin_lock_irqsave(&ctx->lock, flags); |
| 101 | |
| 102 | /* |
| 103 | * The current breakpoint counter is not included in the list |
| 104 | * at the open() callback time |
| 105 | */ |
| 106 | list_for_each_entry(bp, list, event_entry) { |
| 107 | if (bp->attr.type == PERF_TYPE_BREAKPOINT) |
| 108 | count++; |
| 109 | } |
| 110 | |
| 111 | raw_spin_unlock_irqrestore(&ctx->lock, flags); |
| 112 | |
| 113 | return count; |
| 114 | } |
| 115 | |
| 116 | /* |
| 117 | * Report the number of pinned/un-pinned breakpoints we have in |
| 118 | * a given cpu (cpu > -1) or in all of them (cpu = -1). |
| 119 | */ |
| 120 | static void |
| 121 | fetch_bp_busy_slots(struct bp_busy_slots *slots, struct perf_event *bp) |
| 122 | { |
| 123 | int cpu = bp->cpu; |
| 124 | struct task_struct *tsk = bp->ctx->task; |
| 125 | |
| 126 | if (cpu >= 0) { |
| 127 | slots->pinned = per_cpu(nr_cpu_bp_pinned, cpu); |
| 128 | if (!tsk) |
| 129 | slots->pinned += max_task_bp_pinned(cpu); |
| 130 | else |
| 131 | slots->pinned += task_bp_pinned(tsk); |
| 132 | slots->flexible = per_cpu(nr_bp_flexible, cpu); |
| 133 | |
| 134 | return; |
| 135 | } |
| 136 | |
| 137 | for_each_online_cpu(cpu) { |
| 138 | unsigned int nr; |
| 139 | |
| 140 | nr = per_cpu(nr_cpu_bp_pinned, cpu); |
| 141 | if (!tsk) |
| 142 | nr += max_task_bp_pinned(cpu); |
| 143 | else |
| 144 | nr += task_bp_pinned(tsk); |
| 145 | |
| 146 | if (nr > slots->pinned) |
| 147 | slots->pinned = nr; |
| 148 | |
| 149 | nr = per_cpu(nr_bp_flexible, cpu); |
| 150 | |
| 151 | if (nr > slots->flexible) |
| 152 | slots->flexible = nr; |
| 153 | } |
| 154 | } |
| 155 | |
| 156 | /* |
| 157 | * Add a pinned breakpoint for the given task in our constraint table |
| 158 | */ |
| 159 | static void toggle_bp_task_slot(struct task_struct *tsk, int cpu, bool enable) |
| 160 | { |
| 161 | unsigned int *tsk_pinned; |
| 162 | int count = 0; |
| 163 | |
| 164 | count = task_bp_pinned(tsk); |
| 165 | |
| 166 | tsk_pinned = per_cpu(nr_task_bp_pinned, cpu); |
| 167 | if (enable) { |
| 168 | tsk_pinned[count]++; |
| 169 | if (count > 0) |
| 170 | tsk_pinned[count-1]--; |
| 171 | } else { |
| 172 | tsk_pinned[count]--; |
| 173 | if (count > 0) |
| 174 | tsk_pinned[count-1]++; |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | /* |
| 179 | * Add/remove the given breakpoint in our constraint table |
| 180 | */ |
| 181 | static void toggle_bp_slot(struct perf_event *bp, bool enable) |
| 182 | { |
| 183 | int cpu = bp->cpu; |
| 184 | struct task_struct *tsk = bp->ctx->task; |
| 185 | |
| 186 | /* Pinned counter task profiling */ |
| 187 | if (tsk) { |
| 188 | if (cpu >= 0) { |
| 189 | toggle_bp_task_slot(tsk, cpu, enable); |
| 190 | return; |
| 191 | } |
| 192 | |
| 193 | for_each_online_cpu(cpu) |
| 194 | toggle_bp_task_slot(tsk, cpu, enable); |
| 195 | return; |
| 196 | } |
| 197 | |
| 198 | /* Pinned counter cpu profiling */ |
| 199 | if (enable) |
| 200 | per_cpu(nr_cpu_bp_pinned, bp->cpu)++; |
| 201 | else |
| 202 | per_cpu(nr_cpu_bp_pinned, bp->cpu)--; |
| 203 | } |
| 204 | |
| 205 | /* |
| 206 | * Contraints to check before allowing this new breakpoint counter: |
| 207 | * |
| 208 | * == Non-pinned counter == (Considered as pinned for now) |
| 209 | * |
| 210 | * - If attached to a single cpu, check: |
| 211 | * |
| 212 | * (per_cpu(nr_bp_flexible, cpu) || (per_cpu(nr_cpu_bp_pinned, cpu) |
| 213 | * + max(per_cpu(nr_task_bp_pinned, cpu)))) < HBP_NUM |
| 214 | * |
| 215 | * -> If there are already non-pinned counters in this cpu, it means |
| 216 | * there is already a free slot for them. |
| 217 | * Otherwise, we check that the maximum number of per task |
| 218 | * breakpoints (for this cpu) plus the number of per cpu breakpoint |
| 219 | * (for this cpu) doesn't cover every registers. |
| 220 | * |
| 221 | * - If attached to every cpus, check: |
| 222 | * |
| 223 | * (per_cpu(nr_bp_flexible, *) || (max(per_cpu(nr_cpu_bp_pinned, *)) |
| 224 | * + max(per_cpu(nr_task_bp_pinned, *)))) < HBP_NUM |
| 225 | * |
| 226 | * -> This is roughly the same, except we check the number of per cpu |
| 227 | * bp for every cpu and we keep the max one. Same for the per tasks |
| 228 | * breakpoints. |
| 229 | * |
| 230 | * |
| 231 | * == Pinned counter == |
| 232 | * |
| 233 | * - If attached to a single cpu, check: |
| 234 | * |
| 235 | * ((per_cpu(nr_bp_flexible, cpu) > 1) + per_cpu(nr_cpu_bp_pinned, cpu) |
| 236 | * + max(per_cpu(nr_task_bp_pinned, cpu))) < HBP_NUM |
| 237 | * |
| 238 | * -> Same checks as before. But now the nr_bp_flexible, if any, must keep |
| 239 | * one register at least (or they will never be fed). |
| 240 | * |
| 241 | * - If attached to every cpus, check: |
| 242 | * |
| 243 | * ((per_cpu(nr_bp_flexible, *) > 1) + max(per_cpu(nr_cpu_bp_pinned, *)) |
| 244 | * + max(per_cpu(nr_task_bp_pinned, *))) < HBP_NUM |
| 245 | */ |
| 246 | static int __reserve_bp_slot(struct perf_event *bp) |
| 247 | { |
| 248 | struct bp_busy_slots slots = {0}; |
| 249 | |
| 250 | fetch_bp_busy_slots(&slots, bp); |
| 251 | |
| 252 | /* Flexible counters need to keep at least one slot */ |
| 253 | if (slots.pinned + (!!slots.flexible) == HBP_NUM) |
| 254 | return -ENOSPC; |
| 255 | |
| 256 | toggle_bp_slot(bp, true); |
| 257 | |
| 258 | return 0; |
| 259 | } |
| 260 | |
| 261 | int reserve_bp_slot(struct perf_event *bp) |
| 262 | { |
| 263 | int ret; |
| 264 | |
| 265 | mutex_lock(&nr_bp_mutex); |
| 266 | |
| 267 | ret = __reserve_bp_slot(bp); |
| 268 | |
| 269 | mutex_unlock(&nr_bp_mutex); |
| 270 | |
| 271 | return ret; |
| 272 | } |
| 273 | |
| 274 | static void __release_bp_slot(struct perf_event *bp) |
| 275 | { |
| 276 | toggle_bp_slot(bp, false); |
| 277 | } |
| 278 | |
| 279 | void release_bp_slot(struct perf_event *bp) |
| 280 | { |
| 281 | mutex_lock(&nr_bp_mutex); |
| 282 | |
| 283 | __release_bp_slot(bp); |
| 284 | |
| 285 | mutex_unlock(&nr_bp_mutex); |
| 286 | } |
| 287 | |
| 288 | /* |
| 289 | * Allow the kernel debugger to reserve breakpoint slots without |
| 290 | * taking a lock using the dbg_* variant of for the reserve and |
| 291 | * release breakpoint slots. |
| 292 | */ |
| 293 | int dbg_reserve_bp_slot(struct perf_event *bp) |
| 294 | { |
| 295 | if (mutex_is_locked(&nr_bp_mutex)) |
| 296 | return -1; |
| 297 | |
| 298 | return __reserve_bp_slot(bp); |
| 299 | } |
| 300 | |
| 301 | int dbg_release_bp_slot(struct perf_event *bp) |
| 302 | { |
| 303 | if (mutex_is_locked(&nr_bp_mutex)) |
| 304 | return -1; |
| 305 | |
| 306 | __release_bp_slot(bp); |
| 307 | |
| 308 | return 0; |
| 309 | } |
| 310 | |
| 311 | int register_perf_hw_breakpoint(struct perf_event *bp) |
| 312 | { |
| 313 | int ret; |
| 314 | |
| 315 | ret = reserve_bp_slot(bp); |
| 316 | if (ret) |
| 317 | return ret; |
| 318 | |
| 319 | /* |
| 320 | * Ptrace breakpoints can be temporary perf events only |
| 321 | * meant to reserve a slot. In this case, it is created disabled and |
| 322 | * we don't want to check the params right now (as we put a null addr) |
| 323 | * But perf tools create events as disabled and we want to check |
| 324 | * the params for them. |
| 325 | * This is a quick hack that will be removed soon, once we remove |
| 326 | * the tmp breakpoints from ptrace |
| 327 | */ |
| 328 | if (!bp->attr.disabled || !bp->overflow_handler) |
| 329 | ret = arch_validate_hwbkpt_settings(bp, bp->ctx->task); |
| 330 | |
| 331 | /* if arch_validate_hwbkpt_settings() fails then release bp slot */ |
| 332 | if (ret) |
| 333 | release_bp_slot(bp); |
| 334 | |
| 335 | return ret; |
| 336 | } |
| 337 | |
| 338 | /** |
| 339 | * register_user_hw_breakpoint - register a hardware breakpoint for user space |
| 340 | * @attr: breakpoint attributes |
| 341 | * @triggered: callback to trigger when we hit the breakpoint |
| 342 | * @tsk: pointer to 'task_struct' of the process to which the address belongs |
| 343 | */ |
| 344 | struct perf_event * |
| 345 | register_user_hw_breakpoint(struct perf_event_attr *attr, |
| 346 | perf_overflow_handler_t triggered, |
| 347 | struct task_struct *tsk) |
| 348 | { |
| 349 | return perf_event_create_kernel_counter(attr, -1, tsk->pid, triggered); |
| 350 | } |
| 351 | EXPORT_SYMBOL_GPL(register_user_hw_breakpoint); |
| 352 | |
| 353 | /** |
| 354 | * modify_user_hw_breakpoint - modify a user-space hardware breakpoint |
| 355 | * @bp: the breakpoint structure to modify |
| 356 | * @attr: new breakpoint attributes |
| 357 | * @triggered: callback to trigger when we hit the breakpoint |
| 358 | * @tsk: pointer to 'task_struct' of the process to which the address belongs |
| 359 | */ |
| 360 | int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr) |
| 361 | { |
| 362 | u64 old_addr = bp->attr.bp_addr; |
| 363 | u64 old_len = bp->attr.bp_len; |
| 364 | int old_type = bp->attr.bp_type; |
| 365 | int err = 0; |
| 366 | |
| 367 | perf_event_disable(bp); |
| 368 | |
| 369 | bp->attr.bp_addr = attr->bp_addr; |
| 370 | bp->attr.bp_type = attr->bp_type; |
| 371 | bp->attr.bp_len = attr->bp_len; |
| 372 | |
| 373 | if (attr->disabled) |
| 374 | goto end; |
| 375 | |
| 376 | err = arch_validate_hwbkpt_settings(bp, bp->ctx->task); |
| 377 | if (!err) |
| 378 | perf_event_enable(bp); |
| 379 | |
| 380 | if (err) { |
| 381 | bp->attr.bp_addr = old_addr; |
| 382 | bp->attr.bp_type = old_type; |
| 383 | bp->attr.bp_len = old_len; |
| 384 | if (!bp->attr.disabled) |
| 385 | perf_event_enable(bp); |
| 386 | |
| 387 | return err; |
| 388 | } |
| 389 | |
| 390 | end: |
| 391 | bp->attr.disabled = attr->disabled; |
| 392 | |
| 393 | return 0; |
| 394 | } |
| 395 | EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint); |
| 396 | |
| 397 | /** |
| 398 | * unregister_hw_breakpoint - unregister a user-space hardware breakpoint |
| 399 | * @bp: the breakpoint structure to unregister |
| 400 | */ |
| 401 | void unregister_hw_breakpoint(struct perf_event *bp) |
| 402 | { |
| 403 | if (!bp) |
| 404 | return; |
| 405 | perf_event_release_kernel(bp); |
| 406 | } |
| 407 | EXPORT_SYMBOL_GPL(unregister_hw_breakpoint); |
| 408 | |
| 409 | /** |
| 410 | * register_wide_hw_breakpoint - register a wide breakpoint in the kernel |
| 411 | * @attr: breakpoint attributes |
| 412 | * @triggered: callback to trigger when we hit the breakpoint |
| 413 | * |
| 414 | * @return a set of per_cpu pointers to perf events |
| 415 | */ |
| 416 | struct perf_event ** |
| 417 | register_wide_hw_breakpoint(struct perf_event_attr *attr, |
| 418 | perf_overflow_handler_t triggered) |
| 419 | { |
| 420 | struct perf_event **cpu_events, **pevent, *bp; |
| 421 | long err; |
| 422 | int cpu; |
| 423 | |
| 424 | cpu_events = alloc_percpu(typeof(*cpu_events)); |
| 425 | if (!cpu_events) |
| 426 | return ERR_PTR(-ENOMEM); |
| 427 | |
| 428 | get_online_cpus(); |
| 429 | for_each_online_cpu(cpu) { |
| 430 | pevent = per_cpu_ptr(cpu_events, cpu); |
| 431 | bp = perf_event_create_kernel_counter(attr, cpu, -1, triggered); |
| 432 | |
| 433 | *pevent = bp; |
| 434 | |
| 435 | if (IS_ERR(bp)) { |
| 436 | err = PTR_ERR(bp); |
| 437 | goto fail; |
| 438 | } |
| 439 | } |
| 440 | put_online_cpus(); |
| 441 | |
| 442 | return cpu_events; |
| 443 | |
| 444 | fail: |
| 445 | for_each_online_cpu(cpu) { |
| 446 | pevent = per_cpu_ptr(cpu_events, cpu); |
| 447 | if (IS_ERR(*pevent)) |
| 448 | break; |
| 449 | unregister_hw_breakpoint(*pevent); |
| 450 | } |
| 451 | put_online_cpus(); |
| 452 | |
| 453 | free_percpu(cpu_events); |
| 454 | return ERR_PTR(err); |
| 455 | } |
| 456 | EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint); |
| 457 | |
| 458 | /** |
| 459 | * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel |
| 460 | * @cpu_events: the per cpu set of events to unregister |
| 461 | */ |
| 462 | void unregister_wide_hw_breakpoint(struct perf_event **cpu_events) |
| 463 | { |
| 464 | int cpu; |
| 465 | struct perf_event **pevent; |
| 466 | |
| 467 | for_each_possible_cpu(cpu) { |
| 468 | pevent = per_cpu_ptr(cpu_events, cpu); |
| 469 | unregister_hw_breakpoint(*pevent); |
| 470 | } |
| 471 | free_percpu(cpu_events); |
| 472 | } |
| 473 | EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint); |
| 474 | |
| 475 | static struct notifier_block hw_breakpoint_exceptions_nb = { |
| 476 | .notifier_call = hw_breakpoint_exceptions_notify, |
| 477 | /* we need to be notified first */ |
| 478 | .priority = 0x7fffffff |
| 479 | }; |
| 480 | |
| 481 | static int __init init_hw_breakpoint(void) |
| 482 | { |
| 483 | return register_die_notifier(&hw_breakpoint_exceptions_nb); |
| 484 | } |
| 485 | core_initcall(init_hw_breakpoint); |
| 486 | |
| 487 | |
| 488 | struct pmu perf_ops_bp = { |
| 489 | .enable = arch_install_hw_breakpoint, |
| 490 | .disable = arch_uninstall_hw_breakpoint, |
| 491 | .read = hw_breakpoint_pmu_read, |
| 492 | .unthrottle = hw_breakpoint_pmu_unthrottle |
| 493 | }; |