kernel/events/hw_breakpoint.c

   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/slab.h>
  44 #include <linux/list.h>
  45 #include <linux/cpu.h>
  46 #include <linux/smp.h>
  47
  48 #include <linux/hw_breakpoint.h>
  49
  50
  51 /*
  52  * Constraints data
  53  */
  54
  55 /* Number of pinned cpu breakpoints in a cpu */
  56 static DEFINE_PER_CPU(unsigned int, nr_cpu_bp_pinned[TYPE_MAX]);
  57
  58 /* Number of pinned task breakpoints in a cpu */
  59 static DEFINE_PER_CPU(unsigned int *, nr_task_bp_pinned[TYPE_MAX]);
  60
  61 /* Number of non-pinned cpu/task breakpoints in a cpu */
  62 static DEFINE_PER_CPU(unsigned int, nr_bp_flexible[TYPE_MAX]);
  63
  64 static int nr_slots[TYPE_MAX];
  65
  66 /* Keep track of the breakpoints attached to tasks */
  67 static LIST_HEAD(bp_task_head);
  68
  69 static int constraints_initialized;
  70
  71 /* Gather the number of total pinned and un-pinned bp in a cpuset */
  72 struct bp_busy_slots {
  73         unsigned int pinned;
  74         unsigned int flexible;
  75 };
  76
  77 /* Serialize accesses to the above constraints */
  78 static DEFINE_MUTEX(nr_bp_mutex);
  79
  80 __weak int hw_breakpoint_weight(struct perf_event *bp)
  81 {
  82         return 1;
  83 }
  84
  85 static inline enum bp_type_idx find_slot_idx(struct perf_event *bp)
  86 {
  87         if (bp->attr.bp_type & HW_BREAKPOINT_RW)
  88                 return TYPE_DATA;
  89
  90         return TYPE_INST;
  91 }
  92
  93 /*
  94  * Report the maximum number of pinned breakpoints a task
  95  * have in this cpu
  96  */
  97 static unsigned int max_task_bp_pinned(int cpu, enum bp_type_idx type)
  98 {
  99         int i;
 100         unsigned int *tsk_pinned = per_cpu(nr_task_bp_pinned[type], cpu);
 101
 102         for (i = nr_slots[type] - 1; i >= 0; i--) {
 103                 if (tsk_pinned[i] > 0)
 104                         return i + 1;
 105         }
 106
 107         return 0;
 108 }
 109
 110 /*
 111  * Count the number of breakpoints of the same type and same task.
 112  * The given event must be not on the list.
 113  */
 114 static int task_bp_pinned(int cpu, struct perf_event *bp, enum bp_type_idx type)
 115 {
 116         struct task_struct *tsk = bp->hw.bp_target;
 117         struct perf_event *iter;
 118         int count = 0;
 119
 120         list_for_each_entry(iter, &bp_task_head, hw.bp_list) {
 121                 if (iter->hw.bp_target == tsk &&
 122                     find_slot_idx(iter) == type &&
 123                     (iter->cpu < 0 || cpu == iter->cpu))
 124                         count += hw_breakpoint_weight(iter);
 125         }
 126
 127         return count;
 128 }
 129
 130 /*
 131  * Report the number of pinned/un-pinned breakpoints we have in
 132  * a given cpu (cpu > -1) or in all of them (cpu = -1).
 133  */
 134 static void
 135 fetch_bp_busy_slots(struct bp_busy_slots *slots, struct perf_event *bp,
 136                     enum bp_type_idx type)
 137 {
 138         int cpu = bp->cpu;
 139         struct task_struct *tsk = bp->hw.bp_target;
 140
 141         if (cpu >= 0) {
 142                 slots->pinned = per_cpu(nr_cpu_bp_pinned[type], cpu);
 143                 if (!tsk)
 144                         slots->pinned += max_task_bp_pinned(cpu, type);
 145                 else
 146                         slots->pinned += task_bp_pinned(cpu, bp, type);
 147                 slots->flexible = per_cpu(nr_bp_flexible[type], cpu);
 148
 149                 return;
 150         }
 151
 152         for_each_possible_cpu(cpu) {
 153                 unsigned int nr;
 154
 155                 nr = per_cpu(nr_cpu_bp_pinned[type], cpu);
 156                 if (!tsk)
 157                         nr += max_task_bp_pinned(cpu, type);
 158                 else
 159                         nr += task_bp_pinned(cpu, bp, type);
 160
 161                 if (nr > slots->pinned)
 162                         slots->pinned = nr;
 163
 164                 nr = per_cpu(nr_bp_flexible[type], cpu);
 165
 166                 if (nr > slots->flexible)
 167                         slots->flexible = nr;
 168         }
 169 }
 170
 171 /*
 172  * For now, continue to consider flexible as pinned, until we can
 173  * ensure no flexible event can ever be scheduled before a pinned event
 174  * in a same cpu.
 175  */
 176 static void
 177 fetch_this_slot(struct bp_busy_slots *slots, int weight)
 178 {
 179         slots->pinned += weight;
 180 }
 181
 182 /*
 183  * Add a pinned breakpoint for the given task in our constraint table
 184  */
 185 static void toggle_bp_task_slot(struct perf_event *bp, int cpu, bool enable,
 186                                 enum bp_type_idx type, int weight)
 187 {
 188         unsigned int *tsk_pinned;
 189         int old_count = 0;
 190         int old_idx = 0;
 191         int idx = 0;
 192
 193         old_count = task_bp_pinned(cpu, bp, type);
 194         old_idx = old_count - 1;
 195         idx = old_idx + weight;
 196
 197         /* tsk_pinned[n] is the number of tasks having n breakpoints */
 198         tsk_pinned = per_cpu(nr_task_bp_pinned[type], cpu);
 199         if (enable) {
 200                 tsk_pinned[idx]++;
 201                 if (old_count > 0)
 202                         tsk_pinned[old_idx]--;
 203         } else {
 204                 tsk_pinned[idx]--;
 205                 if (old_count > 0)
 206                         tsk_pinned[old_idx]++;
 207         }
 208 }
 209
 210 /*
 211  * Add/remove the given breakpoint in our constraint table
 212  */
 213 static void
 214 toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type,
 215                int weight)
 216 {
 217         int cpu = bp->cpu;
 218         struct task_struct *tsk = bp->hw.bp_target;
 219
 220         /* Pinned counter cpu profiling */
 221         if (!tsk) {
 222
 223                 if (enable)
 224                         per_cpu(nr_cpu_bp_pinned[type], bp->cpu) += weight;
 225                 else
 226                         per_cpu(nr_cpu_bp_pinned[type], bp->cpu) -= weight;
 227                 return;
 228         }
 229
 230         /* Pinned counter task profiling */
 231
 232         if (!enable)
 233                 list_del(&bp->hw.bp_list);
 234
 235         if (cpu >= 0) {
 236                 toggle_bp_task_slot(bp, cpu, enable, type, weight);
 237         } else {
 238                 for_each_possible_cpu(cpu)
 239                         toggle_bp_task_slot(bp, cpu, enable, type, weight);
 240         }
 241
 242         if (enable)
 243                 list_add_tail(&bp->hw.bp_list, &bp_task_head);
 244 }
 245
 246 /*
 247  * Function to perform processor-specific cleanup during unregistration
 248  */
 249 __weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
 250 {
 251         /*
 252          * A weak stub function here for those archs that don't define
 253          * it inside arch/.../kernel/hw_breakpoint.c
 254          */
 255 }
 256
 257 /*
 258  * Contraints to check before allowing this new breakpoint counter:
 259  *
 260  *  == Non-pinned counter == (Considered as pinned for now)
 261  *
 262  *   - If attached to a single cpu, check:
 263  *
 264  *       (per_cpu(nr_bp_flexible, cpu) || (per_cpu(nr_cpu_bp_pinned, cpu)
 265  *           + max(per_cpu(nr_task_bp_pinned, cpu)))) < HBP_NUM
 266  *
 267  *       -> If there are already non-pinned counters in this cpu, it means
 268  *          there is already a free slot for them.
 269  *          Otherwise, we check that the maximum number of per task
 270  *          breakpoints (for this cpu) plus the number of per cpu breakpoint
 271  *          (for this cpu) doesn't cover every registers.
 272  *
 273  *   - If attached to every cpus, check:
 274  *
 275  *       (per_cpu(nr_bp_flexible, *) || (max(per_cpu(nr_cpu_bp_pinned, *))
 276  *           + max(per_cpu(nr_task_bp_pinned, *)))) < HBP_NUM
 277  *
 278  *       -> This is roughly the same, except we check the number of per cpu
 279  *          bp for every cpu and we keep the max one. Same for the per tasks
 280  *          breakpoints.
 281  *
 282  *
 283  * == Pinned counter ==
 284  *
 285  *   - If attached to a single cpu, check:
 286  *
 287  *       ((per_cpu(nr_bp_flexible, cpu) > 1) + per_cpu(nr_cpu_bp_pinned, cpu)
 288  *            + max(per_cpu(nr_task_bp_pinned, cpu))) < HBP_NUM
 289  *
 290  *       -> Same checks as before. But now the nr_bp_flexible, if any, must keep
 291  *          one register at least (or they will never be fed).
 292  *
 293  *   - If attached to every cpus, check:
 294  *
 295  *       ((per_cpu(nr_bp_flexible, *) > 1) + max(per_cpu(nr_cpu_bp_pinned, *))
 296  *            + max(per_cpu(nr_task_bp_pinned, *))) < HBP_NUM
 297  */
 298 static int __reserve_bp_slot(struct perf_event *bp)
 299 {
 300         struct bp_busy_slots slots = {0};
 301         enum bp_type_idx type;
 302         int weight;
 303
 304         /* We couldn't initialize breakpoint constraints on boot */
 305         if (!constraints_initialized)
 306                 return -ENOMEM;
 307
 308         /* Basic checks */
 309         if (bp->attr.bp_type == HW_BREAKPOINT_EMPTY ||
 310             bp->attr.bp_type == HW_BREAKPOINT_INVALID)
 311                 return -EINVAL;
 312
 313         type = find_slot_idx(bp);
 314         weight = hw_breakpoint_weight(bp);
 315
 316         fetch_bp_busy_slots(&slots, bp, type);
 317         /*
 318          * Simulate the addition of this breakpoint to the constraints
 319          * and see the result.
 320          */
 321         fetch_this_slot(&slots, weight);
 322
 323         /* Flexible counters need to keep at least one slot */
 324         if (slots.pinned + (!!slots.flexible) > nr_slots[type])
 325                 return -ENOSPC;
 326
 327         toggle_bp_slot(bp, true, type, weight);
 328
 329         return 0;
 330 }
 331
 332 int reserve_bp_slot(struct perf_event *bp)
 333 {
 334         int ret;
 335
 336         mutex_lock(&nr_bp_mutex);
 337
 338         ret = __reserve_bp_slot(bp);
 339
 340         mutex_unlock(&nr_bp_mutex);
 341
 342         return ret;
 343 }
 344
 345 static void __release_bp_slot(struct perf_event *bp)
 346 {
 347         enum bp_type_idx type;
 348         int weight;
 349
 350         type = find_slot_idx(bp);
 351         weight = hw_breakpoint_weight(bp);
 352         toggle_bp_slot(bp, false, type, weight);
 353 }
 354
 355 void release_bp_slot(struct perf_event *bp)
 356 {
 357         mutex_lock(&nr_bp_mutex);
 358
 359         arch_unregister_hw_breakpoint(bp);
 360         __release_bp_slot(bp);
 361
 362         mutex_unlock(&nr_bp_mutex);
 363 }
 364
 365 /*
 366  * Allow the kernel debugger to reserve breakpoint slots without
 367  * taking a lock using the dbg_* variant of for the reserve and
 368  * release breakpoint slots.
 369  */
 370 int dbg_reserve_bp_slot(struct perf_event *bp)
 371 {
 372         if (mutex_is_locked(&nr_bp_mutex))
 373                 return -1;
 374
 375         return __reserve_bp_slot(bp);
 376 }
 377
 378 int dbg_release_bp_slot(struct perf_event *bp)
 379 {
 380         if (mutex_is_locked(&nr_bp_mutex))
 381                 return -1;
 382
 383         __release_bp_slot(bp);
 384
 385         return 0;
 386 }
 387
 388 static int validate_hw_breakpoint(struct perf_event *bp)
 389 {
 390         int ret;
 391
 392         ret = arch_validate_hwbkpt_settings(bp);
 393         if (ret)
 394                 return ret;
 395
 396         if (arch_check_bp_in_kernelspace(bp)) {
 397                 if (bp->attr.exclude_kernel)
 398                         return -EINVAL;
 399                 /*
 400                  * Don't let unprivileged users set a breakpoint in the trap
 401                  * path to avoid trap recursion attacks.
 402                  */
 403                 if (!capable(CAP_SYS_ADMIN))
 404                         return -EPERM;
 405         }
 406
 407         return 0;
 408 }
 409
 410 int register_perf_hw_breakpoint(struct perf_event *bp)
 411 {
 412         int ret;
 413
 414         ret = reserve_bp_slot(bp);
 415         if (ret)
 416                 return ret;
 417
 418         ret = validate_hw_breakpoint(bp);
 419
 420         /* if arch_validate_hwbkpt_settings() fails then release bp slot */
 421         if (ret)
 422                 release_bp_slot(bp);
 423
 424         return ret;
 425 }
 426
 427 /**
 428  * register_user_hw_breakpoint - register a hardware breakpoint for user space
 429  * @attr: breakpoint attributes
 430  * @triggered: callback to trigger when we hit the breakpoint
 431  * @tsk: pointer to 'task_struct' of the process to which the address belongs
 432  */
 433 struct perf_event *
 434 register_user_hw_breakpoint(struct perf_event_attr *attr,
 435                             perf_overflow_handler_t triggered,
 436                             void *context,
 437                             struct task_struct *tsk)
 438 {
 439         return perf_event_create_kernel_counter(attr, -1, tsk, triggered,
 440                                                 context);
 441 }
 442 EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
 443
 444 /**
 445  * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
 446  * @bp: the breakpoint structure to modify
 447  * @attr: new breakpoint attributes
 448  * @triggered: callback to trigger when we hit the breakpoint
 449  * @tsk: pointer to 'task_struct' of the process to which the address belongs
 450  */
 451 int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
 452 {
 453         u64 old_addr = bp->attr.bp_addr;
 454         u64 old_len = bp->attr.bp_len;
 455         int old_type = bp->attr.bp_type;
 456         int err = 0;
 457
 458         /*
 459          * modify_user_hw_breakpoint can be invoked with IRQs disabled and hence it
 460          * will not be possible to raise IPIs that invoke __perf_event_disable.
 461          * So call the function directly after making sure we are targeting the
 462          * current task.
 463          */
 464         if (irqs_disabled() && bp->ctx && bp->ctx->task == current)
 465                 __perf_event_disable(bp);
 466         else
 467                 perf_event_disable(bp);
 468
 469         bp->attr.bp_addr = attr->bp_addr;
 470         bp->attr.bp_type = attr->bp_type;
 471         bp->attr.bp_len = attr->bp_len;
 472
 473         if (attr->disabled)
 474                 goto end;
 475
 476         err = validate_hw_breakpoint(bp);
 477         if (!err)
 478                 perf_event_enable(bp);
 479
 480         if (err) {
 481                 bp->attr.bp_addr = old_addr;
 482                 bp->attr.bp_type = old_type;
 483                 bp->attr.bp_len = old_len;
 484                 if (!bp->attr.disabled)
 485                         perf_event_enable(bp);
 486
 487                 return err;
 488         }
 489
 490 end:
 491         bp->attr.disabled = attr->disabled;
 492
 493         return 0;
 494 }
 495 EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
 496
 497 /**
 498  * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
 499  * @bp: the breakpoint structure to unregister
 500  */
 501 void unregister_hw_breakpoint(struct perf_event *bp)
 502 {
 503         if (!bp)
 504                 return;
 505         perf_event_release_kernel(bp);
 506 }
 507 EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
 508
 509 /**
 510  * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
 511  * @attr: breakpoint attributes
 512  * @triggered: callback to trigger when we hit the breakpoint
 513  *
 514  * @return a set of per_cpu pointers to perf events
 515  */
 516 struct perf_event * __percpu *
 517 register_wide_hw_breakpoint(struct perf_event_attr *attr,
 518                             perf_overflow_handler_t triggered,
 519                             void *context)
 520 {
 521         struct perf_event * __percpu *cpu_events, **pevent, *bp;
 522         long err;
 523         int cpu;
 524
 525         cpu_events = alloc_percpu(typeof(*cpu_events));
 526         if (!cpu_events)
 527                 return (void __percpu __force *)ERR_PTR(-ENOMEM);
 528
 529         get_online_cpus();
 530         for_each_online_cpu(cpu) {
 531                 pevent = per_cpu_ptr(cpu_events, cpu);
 532                 bp = perf_event_create_kernel_counter(attr, cpu, NULL,
 533                                                       triggered, context);
 534
 535                 *pevent = bp;
 536
 537                 if (IS_ERR(bp)) {
 538                         err = PTR_ERR(bp);
 539                         goto fail;
 540                 }
 541         }
 542         put_online_cpus();
 543
 544         return cpu_events;
 545
 546 fail:
 547         for_each_online_cpu(cpu) {
 548                 pevent = per_cpu_ptr(cpu_events, cpu);
 549                 if (IS_ERR(*pevent))
 550                         break;
 551                 unregister_hw_breakpoint(*pevent);
 552         }
 553         put_online_cpus();
 554
 555         free_percpu(cpu_events);
 556         return (void __percpu __force *)ERR_PTR(err);
 557 }
 558 EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
 559
 560 /**
 561  * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
 562  * @cpu_events: the per cpu set of events to unregister
 563  */
 564 void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
 565 {
 566         int cpu;
 567         struct perf_event **pevent;
 568
 569         for_each_possible_cpu(cpu) {
 570                 pevent = per_cpu_ptr(cpu_events, cpu);
 571                 unregister_hw_breakpoint(*pevent);
 572         }
 573         free_percpu(cpu_events);
 574 }
 575 EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
 576
 577 static struct notifier_block hw_breakpoint_exceptions_nb = {
 578         .notifier_call = hw_breakpoint_exceptions_notify,
 579         /* we need to be notified first */
 580         .priority = 0x7fffffff
 581 };
 582
 583 static void bp_perf_event_destroy(struct perf_event *event)
 584 {
 585         release_bp_slot(event);
 586 }
 587
 588 static int hw_breakpoint_event_init(struct perf_event *bp)
 589 {
 590         int err;
 591
 592         if (bp->attr.type != PERF_TYPE_BREAKPOINT)
 593                 return -ENOENT;
 594
 595         /*
 596          * no branch sampling for breakpoint events
 597          */
 598         if (has_branch_stack(bp))
 599                 return -EOPNOTSUPP;
 600
 601         err = register_perf_hw_breakpoint(bp);
 602         if (err)
 603                 return err;
 604
 605         bp->destroy = bp_perf_event_destroy;
 606
 607         return 0;
 608 }
 609
 610 static int hw_breakpoint_add(struct perf_event *bp, int flags)
 611 {
 612         if (!(flags & PERF_EF_START))
 613                 bp->hw.state = PERF_HES_STOPPED;
 614
 615         return arch_install_hw_breakpoint(bp);
 616 }
 617
 618 static void hw_breakpoint_del(struct perf_event *bp, int flags)
 619 {
 620         arch_uninstall_hw_breakpoint(bp);
 621 }
 622
 623 static void hw_breakpoint_start(struct perf_event *bp, int flags)
 624 {
 625         bp->hw.state = 0;
 626 }
 627
 628 static void hw_breakpoint_stop(struct perf_event *bp, int flags)
 629 {
 630         bp->hw.state = PERF_HES_STOPPED;
 631 }
 632
 633 static int hw_breakpoint_event_idx(struct perf_event *bp)
 634 {
 635         return 0;
 636 }
 637
 638 static struct pmu perf_breakpoint = {
 639         .task_ctx_nr    = perf_sw_context, /* could eventually get its own */
 640
 641         .event_init     = hw_breakpoint_event_init,
 642         .add            = hw_breakpoint_add,
 643         .del            = hw_breakpoint_del,
 644         .start          = hw_breakpoint_start,
 645         .stop           = hw_breakpoint_stop,
 646         .read           = hw_breakpoint_pmu_read,
 647
 648         .event_idx      = hw_breakpoint_event_idx,
 649 };
 650
 651 int __init init_hw_breakpoint(void)
 652 {
 653         unsigned int **task_bp_pinned;
 654         int cpu, err_cpu;
 655         int i;
 656
 657         for (i = 0; i < TYPE_MAX; i++)
 658                 nr_slots[i] = hw_breakpoint_slots(i);
 659
 660         for_each_possible_cpu(cpu) {
 661                 for (i = 0; i < TYPE_MAX; i++) {
 662                         task_bp_pinned = &per_cpu(nr_task_bp_pinned[i], cpu);
 663                         *task_bp_pinned = kzalloc(sizeof(int) * nr_slots[i],
 664                                                   GFP_KERNEL);
 665                         if (!*task_bp_pinned)
 666                                 goto err_alloc;
 667                 }
 668         }
 669
 670         constraints_initialized = 1;
 671
 672         perf_pmu_register(&perf_breakpoint, "breakpoint", PERF_TYPE_BREAKPOINT);
 673
 674         return register_die_notifier(&hw_breakpoint_exceptions_nb);
 675
 676  err_alloc:
 677         for_each_possible_cpu(err_cpu) {
 678                 for (i = 0; i < TYPE_MAX; i++)
 679                         kfree(per_cpu(nr_task_bp_pinned[i], err_cpu));
 680                 if (err_cpu == cpu)
 681                         break;
 682         }
 683
 684         return -ENOMEM;
 685 }
 686
 687