kernel/events/hw_breakpoint.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Copyright (C) 2007 Alan Stern
   4  * Copyright (C) IBM Corporation, 2009
   5  * Copyright (C) 2009, Frederic Weisbecker <fweisbec@gmail.com>
   6  *
   7  * Thanks to Ingo Molnar for his many suggestions.
   8  *
   9  * Authors: Alan Stern <stern@rowland.harvard.edu>
  10  *          K.Prasad <prasad@linux.vnet.ibm.com>
  11  *          Frederic Weisbecker <fweisbec@gmail.com>
  12  */
  13
  14 /*
  15  * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
  16  * using the CPU's debug registers.
  17  * This file contains the arch-independent routines.
  18  */
  19
  20 #include <linux/irqflags.h>
  21 #include <linux/kallsyms.h>
  22 #include <linux/notifier.h>
  23 #include <linux/kprobes.h>
  24 #include <linux/kdebug.h>
  25 #include <linux/kernel.h>
  26 #include <linux/module.h>
  27 #include <linux/percpu.h>
  28 #include <linux/sched.h>
  29 #include <linux/init.h>
  30 #include <linux/slab.h>
  31 #include <linux/list.h>
  32 #include <linux/cpu.h>
  33 #include <linux/smp.h>
  34 #include <linux/bug.h>
  35
  36 #include <linux/hw_breakpoint.h>
  37 /*
  38  * Constraints data
  39  */
  40 struct bp_cpuinfo {
  41         /* Number of pinned cpu breakpoints in a cpu */
  42         unsigned int    cpu_pinned;
  43         /* tsk_pinned[n] is the number of tasks having n+1 breakpoints */
  44         unsigned int    *tsk_pinned;
  45         /* Number of non-pinned cpu/task breakpoints in a cpu */
  46         unsigned int    flexible; /* XXX: placeholder, see fetch_this_slot() */
  47 };
  48
  49 static DEFINE_PER_CPU(struct bp_cpuinfo, bp_cpuinfo[TYPE_MAX]);
  50 static int nr_slots[TYPE_MAX];
  51
  52 static struct bp_cpuinfo *get_bp_info(int cpu, enum bp_type_idx type)
  53 {
  54         return per_cpu_ptr(bp_cpuinfo + type, cpu);
  55 }
  56
  57 /* Keep track of the breakpoints attached to tasks */
  58 static LIST_HEAD(bp_task_head);
  59
  60 static int constraints_initialized;
  61
  62 /* Gather the number of total pinned and un-pinned bp in a cpuset */
  63 struct bp_busy_slots {
  64         unsigned int pinned;
  65         unsigned int flexible;
  66 };
  67
  68 /* Serialize accesses to the above constraints */
  69 static DEFINE_MUTEX(nr_bp_mutex);
  70
  71 __weak int hw_breakpoint_weight(struct perf_event *bp)
  72 {
  73         return 1;
  74 }
  75
  76 static inline enum bp_type_idx find_slot_idx(u64 bp_type)
  77 {
  78         if (bp_type & HW_BREAKPOINT_RW)
  79                 return TYPE_DATA;
  80
  81         return TYPE_INST;
  82 }
  83
  84 /*
  85  * Report the maximum number of pinned breakpoints a task
  86  * have in this cpu
  87  */
  88 static unsigned int max_task_bp_pinned(int cpu, enum bp_type_idx type)
  89 {
  90         unsigned int *tsk_pinned = get_bp_info(cpu, type)->tsk_pinned;
  91         int i;
  92
  93         for (i = nr_slots[type] - 1; i >= 0; i--) {
  94                 if (tsk_pinned[i] > 0)
  95                         return i + 1;
  96         }
  97
  98         return 0;
  99 }
 100
 101 /*
 102  * Count the number of breakpoints of the same type and same task.
 103  * The given event must be not on the list.
 104  */
 105 static int task_bp_pinned(int cpu, struct perf_event *bp, enum bp_type_idx type)
 106 {
 107         struct task_struct *tsk = bp->hw.target;
 108         struct perf_event *iter;
 109         int count = 0;
 110
 111         list_for_each_entry(iter, &bp_task_head, hw.bp_list) {
 112                 if (iter->hw.target == tsk &&
 113                     find_slot_idx(iter->attr.bp_type) == type &&
 114                     (iter->cpu < 0 || cpu == iter->cpu))
 115                         count += hw_breakpoint_weight(iter);
 116         }
 117
 118         return count;
 119 }
 120
 121 static const struct cpumask *cpumask_of_bp(struct perf_event *bp)
 122 {
 123         if (bp->cpu >= 0)
 124                 return cpumask_of(bp->cpu);
 125         return cpu_possible_mask;
 126 }
 127
 128 /*
 129  * Report the number of pinned/un-pinned breakpoints we have in
 130  * a given cpu (cpu > -1) or in all of them (cpu = -1).
 131  */
 132 static void
 133 fetch_bp_busy_slots(struct bp_busy_slots *slots, struct perf_event *bp,
 134                     enum bp_type_idx type)
 135 {
 136         const struct cpumask *cpumask = cpumask_of_bp(bp);
 137         int cpu;
 138
 139         for_each_cpu(cpu, cpumask) {
 140                 struct bp_cpuinfo *info = get_bp_info(cpu, type);
 141                 int nr;
 142
 143                 nr = info->cpu_pinned;
 144                 if (!bp->hw.target)
 145                         nr += max_task_bp_pinned(cpu, type);
 146                 else
 147                         nr += task_bp_pinned(cpu, bp, type);
 148
 149                 if (nr > slots->pinned)
 150                         slots->pinned = nr;
 151
 152                 nr = info->flexible;
 153                 if (nr > slots->flexible)
 154                         slots->flexible = nr;
 155         }
 156 }
 157
 158 /*
 159  * For now, continue to consider flexible as pinned, until we can
 160  * ensure no flexible event can ever be scheduled before a pinned event
 161  * in a same cpu.
 162  */
 163 static void
 164 fetch_this_slot(struct bp_busy_slots *slots, int weight)
 165 {
 166         slots->pinned += weight;
 167 }
 168
 169 /*
 170  * Add a pinned breakpoint for the given task in our constraint table
 171  */
 172 static void toggle_bp_task_slot(struct perf_event *bp, int cpu,
 173                                 enum bp_type_idx type, int weight)
 174 {
 175         unsigned int *tsk_pinned = get_bp_info(cpu, type)->tsk_pinned;
 176         int old_idx, new_idx;
 177
 178         old_idx = task_bp_pinned(cpu, bp, type) - 1;
 179         new_idx = old_idx + weight;
 180
 181         if (old_idx >= 0)
 182                 tsk_pinned[old_idx]--;
 183         if (new_idx >= 0)
 184                 tsk_pinned[new_idx]++;
 185 }
 186
 187 /*
 188  * Add/remove the given breakpoint in our constraint table
 189  */
 190 static void
 191 toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type,
 192                int weight)
 193 {
 194         const struct cpumask *cpumask = cpumask_of_bp(bp);
 195         int cpu;
 196
 197         if (!enable)
 198                 weight = -weight;
 199
 200         /* Pinned counter cpu profiling */
 201         if (!bp->hw.target) {
 202                 get_bp_info(bp->cpu, type)->cpu_pinned += weight;
 203                 return;
 204         }
 205
 206         /* Pinned counter task profiling */
 207         for_each_cpu(cpu, cpumask)
 208                 toggle_bp_task_slot(bp, cpu, type, weight);
 209
 210         if (enable)
 211                 list_add_tail(&bp->hw.bp_list, &bp_task_head);
 212         else
 213                 list_del(&bp->hw.bp_list);
 214 }
 215
 216 /*
 217  * Function to perform processor-specific cleanup during unregistration
 218  */
 219 __weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
 220 {
 221         /*
 222          * A weak stub function here for those archs that don't define
 223          * it inside arch/.../kernel/hw_breakpoint.c
 224          */
 225 }
 226
 227 /*
 228  * Constraints to check before allowing this new breakpoint counter:
 229  *
 230  *  == Non-pinned counter == (Considered as pinned for now)
 231  *
 232  *   - If attached to a single cpu, check:
 233  *
 234  *       (per_cpu(info->flexible, cpu) || (per_cpu(info->cpu_pinned, cpu)
 235  *           + max(per_cpu(info->tsk_pinned, cpu)))) < HBP_NUM
 236  *
 237  *       -> If there are already non-pinned counters in this cpu, it means
 238  *          there is already a free slot for them.
 239  *          Otherwise, we check that the maximum number of per task
 240  *          breakpoints (for this cpu) plus the number of per cpu breakpoint
 241  *          (for this cpu) doesn't cover every registers.
 242  *
 243  *   - If attached to every cpus, check:
 244  *
 245  *       (per_cpu(info->flexible, *) || (max(per_cpu(info->cpu_pinned, *))
 246  *           + max(per_cpu(info->tsk_pinned, *)))) < HBP_NUM
 247  *
 248  *       -> This is roughly the same, except we check the number of per cpu
 249  *          bp for every cpu and we keep the max one. Same for the per tasks
 250  *          breakpoints.
 251  *
 252  *
 253  * == Pinned counter ==
 254  *
 255  *   - If attached to a single cpu, check:
 256  *
 257  *       ((per_cpu(info->flexible, cpu) > 1) + per_cpu(info->cpu_pinned, cpu)
 258  *            + max(per_cpu(info->tsk_pinned, cpu))) < HBP_NUM
 259  *
 260  *       -> Same checks as before. But now the info->flexible, if any, must keep
 261  *          one register at least (or they will never be fed).
 262  *
 263  *   - If attached to every cpus, check:
 264  *
 265  *       ((per_cpu(info->flexible, *) > 1) + max(per_cpu(info->cpu_pinned, *))
 266  *            + max(per_cpu(info->tsk_pinned, *))) < HBP_NUM
 267  */
 268 static int __reserve_bp_slot(struct perf_event *bp, u64 bp_type)
 269 {
 270         struct bp_busy_slots slots = {0};
 271         enum bp_type_idx type;
 272         int weight;
 273
 274         /* We couldn't initialize breakpoint constraints on boot */
 275         if (!constraints_initialized)
 276                 return -ENOMEM;
 277
 278         /* Basic checks */
 279         if (bp_type == HW_BREAKPOINT_EMPTY ||
 280             bp_type == HW_BREAKPOINT_INVALID)
 281                 return -EINVAL;
 282
 283         type = find_slot_idx(bp_type);
 284         weight = hw_breakpoint_weight(bp);
 285
 286         fetch_bp_busy_slots(&slots, bp, type);
 287         /*
 288          * Simulate the addition of this breakpoint to the constraints
 289          * and see the result.
 290          */
 291         fetch_this_slot(&slots, weight);
 292
 293         /* Flexible counters need to keep at least one slot */
 294         if (slots.pinned + (!!slots.flexible) > nr_slots[type])
 295                 return -ENOSPC;
 296
 297         toggle_bp_slot(bp, true, type, weight);
 298
 299         return 0;
 300 }
 301
 302 int reserve_bp_slot(struct perf_event *bp)
 303 {
 304         int ret;
 305
 306         mutex_lock(&nr_bp_mutex);
 307
 308         ret = __reserve_bp_slot(bp, bp->attr.bp_type);
 309
 310         mutex_unlock(&nr_bp_mutex);
 311
 312         return ret;
 313 }
 314
 315 static void __release_bp_slot(struct perf_event *bp, u64 bp_type)
 316 {
 317         enum bp_type_idx type;
 318         int weight;
 319
 320         type = find_slot_idx(bp_type);
 321         weight = hw_breakpoint_weight(bp);
 322         toggle_bp_slot(bp, false, type, weight);
 323 }
 324
 325 void release_bp_slot(struct perf_event *bp)
 326 {
 327         mutex_lock(&nr_bp_mutex);
 328
 329         arch_unregister_hw_breakpoint(bp);
 330         __release_bp_slot(bp, bp->attr.bp_type);
 331
 332         mutex_unlock(&nr_bp_mutex);
 333 }
 334
 335 static int __modify_bp_slot(struct perf_event *bp, u64 old_type, u64 new_type)
 336 {
 337         int err;
 338
 339         __release_bp_slot(bp, old_type);
 340
 341         err = __reserve_bp_slot(bp, new_type);
 342         if (err) {
 343                 /*
 344                  * Reserve the old_type slot back in case
 345                  * there's no space for the new type.
 346                  *
 347                  * This must succeed, because we just released
 348                  * the old_type slot in the __release_bp_slot
 349                  * call above. If not, something is broken.
 350                  */
 351                 WARN_ON(__reserve_bp_slot(bp, old_type));
 352         }
 353
 354         return err;
 355 }
 356
 357 static int modify_bp_slot(struct perf_event *bp, u64 old_type, u64 new_type)
 358 {
 359         int ret;
 360
 361         mutex_lock(&nr_bp_mutex);
 362         ret = __modify_bp_slot(bp, old_type, new_type);
 363         mutex_unlock(&nr_bp_mutex);
 364         return ret;
 365 }
 366
 367 /*
 368  * Allow the kernel debugger to reserve breakpoint slots without
 369  * taking a lock using the dbg_* variant of for the reserve and
 370  * release breakpoint slots.
 371  */
 372 int dbg_reserve_bp_slot(struct perf_event *bp)
 373 {
 374         if (mutex_is_locked(&nr_bp_mutex))
 375                 return -1;
 376
 377         return __reserve_bp_slot(bp, bp->attr.bp_type);
 378 }
 379
 380 int dbg_release_bp_slot(struct perf_event *bp)
 381 {
 382         if (mutex_is_locked(&nr_bp_mutex))
 383                 return -1;
 384
 385         __release_bp_slot(bp, bp->attr.bp_type);
 386
 387         return 0;
 388 }
 389
 390 static int hw_breakpoint_parse(struct perf_event *bp,
 391                                const struct perf_event_attr *attr,
 392                                struct arch_hw_breakpoint *hw)
 393 {
 394         int err;
 395
 396         err = hw_breakpoint_arch_parse(bp, attr, hw);
 397         if (err)
 398                 return err;
 399
 400         if (arch_check_bp_in_kernelspace(hw)) {
 401                 if (attr->exclude_kernel)
 402                         return -EINVAL;
 403                 /*
 404                  * Don't let unprivileged users set a breakpoint in the trap
 405                  * path to avoid trap recursion attacks.
 406                  */
 407                 if (!capable(CAP_SYS_ADMIN))
 408                         return -EPERM;
 409         }
 410
 411         return 0;
 412 }
 413
 414 int register_perf_hw_breakpoint(struct perf_event *bp)
 415 {
 416         struct arch_hw_breakpoint hw;
 417         int err;
 418
 419         err = reserve_bp_slot(bp);
 420         if (err)
 421                 return err;
 422
 423         err = hw_breakpoint_parse(bp, &bp->attr, &hw);
 424         if (err) {
 425                 release_bp_slot(bp);
 426                 return err;
 427         }
 428
 429         bp->hw.info = hw;
 430
 431         return 0;
 432 }
 433
 434 /**
 435  * register_user_hw_breakpoint - register a hardware breakpoint for user space
 436  * @attr: breakpoint attributes
 437  * @triggered: callback to trigger when we hit the breakpoint
 438  * @tsk: pointer to 'task_struct' of the process to which the address belongs
 439  */
 440 struct perf_event *
 441 register_user_hw_breakpoint(struct perf_event_attr *attr,
 442                             perf_overflow_handler_t triggered,
 443                             void *context,
 444                             struct task_struct *tsk)
 445 {
 446         return perf_event_create_kernel_counter(attr, -1, tsk, triggered,
 447                                                 context);
 448 }
 449 EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
 450
 451 static void hw_breakpoint_copy_attr(struct perf_event_attr *to,
 452                                     struct perf_event_attr *from)
 453 {
 454         to->bp_addr = from->bp_addr;
 455         to->bp_type = from->bp_type;
 456         to->bp_len  = from->bp_len;
 457         to->disabled = from->disabled;
 458 }
 459
 460 int
 461 modify_user_hw_breakpoint_check(struct perf_event *bp, struct perf_event_attr *attr,
 462                                 bool check)
 463 {
 464         struct arch_hw_breakpoint hw;
 465         int err;
 466
 467         err = hw_breakpoint_parse(bp, attr, &hw);
 468         if (err)
 469                 return err;
 470
 471         if (check) {
 472                 struct perf_event_attr old_attr;
 473
 474                 old_attr = bp->attr;
 475                 hw_breakpoint_copy_attr(&old_attr, attr);
 476                 if (memcmp(&old_attr, attr, sizeof(*attr)))
 477                         return -EINVAL;
 478         }
 479
 480         if (bp->attr.bp_type != attr->bp_type) {
 481                 err = modify_bp_slot(bp, bp->attr.bp_type, attr->bp_type);
 482                 if (err)
 483                         return err;
 484         }
 485
 486         hw_breakpoint_copy_attr(&bp->attr, attr);
 487         bp->hw.info = hw;
 488
 489         return 0;
 490 }
 491
 492 /**
 493  * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
 494  * @bp: the breakpoint structure to modify
 495  * @attr: new breakpoint attributes
 496  */
 497 int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
 498 {
 499         int err;
 500
 501         /*
 502          * modify_user_hw_breakpoint can be invoked with IRQs disabled and hence it
 503          * will not be possible to raise IPIs that invoke __perf_event_disable.
 504          * So call the function directly after making sure we are targeting the
 505          * current task.
 506          */
 507         if (irqs_disabled() && bp->ctx && bp->ctx->task == current)
 508                 perf_event_disable_local(bp);
 509         else
 510                 perf_event_disable(bp);
 511
 512         err = modify_user_hw_breakpoint_check(bp, attr, false);
 513
 514         if (!bp->attr.disabled)
 515                 perf_event_enable(bp);
 516
 517         return err;
 518 }
 519 EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
 520
 521 /**
 522  * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
 523  * @bp: the breakpoint structure to unregister
 524  */
 525 void unregister_hw_breakpoint(struct perf_event *bp)
 526 {
 527         if (!bp)
 528                 return;
 529         perf_event_release_kernel(bp);
 530 }
 531 EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
 532
 533 /**
 534  * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
 535  * @attr: breakpoint attributes
 536  * @triggered: callback to trigger when we hit the breakpoint
 537  *
 538  * @return a set of per_cpu pointers to perf events
 539  */
 540 struct perf_event * __percpu *
 541 register_wide_hw_breakpoint(struct perf_event_attr *attr,
 542                             perf_overflow_handler_t triggered,
 543                             void *context)
 544 {
 545         struct perf_event * __percpu *cpu_events, *bp;
 546         long err = 0;
 547         int cpu;
 548
 549         cpu_events = alloc_percpu(typeof(*cpu_events));
 550         if (!cpu_events)
 551                 return (void __percpu __force *)ERR_PTR(-ENOMEM);
 552
 553         get_online_cpus();
 554         for_each_online_cpu(cpu) {
 555                 bp = perf_event_create_kernel_counter(attr, cpu, NULL,
 556                                                       triggered, context);
 557                 if (IS_ERR(bp)) {
 558                         err = PTR_ERR(bp);
 559                         break;
 560                 }
 561
 562                 per_cpu(*cpu_events, cpu) = bp;
 563         }
 564         put_online_cpus();
 565
 566         if (likely(!err))
 567                 return cpu_events;
 568
 569         unregister_wide_hw_breakpoint(cpu_events);
 570         return (void __percpu __force *)ERR_PTR(err);
 571 }
 572 EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
 573
 574 /**
 575  * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
 576  * @cpu_events: the per cpu set of events to unregister
 577  */
 578 void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
 579 {
 580         int cpu;
 581
 582         for_each_possible_cpu(cpu)
 583                 unregister_hw_breakpoint(per_cpu(*cpu_events, cpu));
 584
 585         free_percpu(cpu_events);
 586 }
 587 EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
 588
 589 static struct notifier_block hw_breakpoint_exceptions_nb = {
 590         .notifier_call = hw_breakpoint_exceptions_notify,
 591         /* we need to be notified first */
 592         .priority = 0x7fffffff
 593 };
 594
 595 static void bp_perf_event_destroy(struct perf_event *event)
 596 {
 597         release_bp_slot(event);
 598 }
 599
 600 static int hw_breakpoint_event_init(struct perf_event *bp)
 601 {
 602         int err;
 603
 604         if (bp->attr.type != PERF_TYPE_BREAKPOINT)
 605                 return -ENOENT;
 606
 607         /*
 608          * no branch sampling for breakpoint events
 609          */
 610         if (has_branch_stack(bp))
 611                 return -EOPNOTSUPP;
 612
 613         err = register_perf_hw_breakpoint(bp);
 614         if (err)
 615                 return err;
 616
 617         bp->destroy = bp_perf_event_destroy;
 618
 619         return 0;
 620 }
 621
 622 static int hw_breakpoint_add(struct perf_event *bp, int flags)
 623 {
 624         if (!(flags & PERF_EF_START))
 625                 bp->hw.state = PERF_HES_STOPPED;
 626
 627         if (is_sampling_event(bp)) {
 628                 bp->hw.last_period = bp->hw.sample_period;
 629                 perf_swevent_set_period(bp);
 630         }
 631
 632         return arch_install_hw_breakpoint(bp);
 633 }
 634
 635 static void hw_breakpoint_del(struct perf_event *bp, int flags)
 636 {
 637         arch_uninstall_hw_breakpoint(bp);
 638 }
 639
 640 static void hw_breakpoint_start(struct perf_event *bp, int flags)
 641 {
 642         bp->hw.state = 0;
 643 }
 644
 645 static void hw_breakpoint_stop(struct perf_event *bp, int flags)
 646 {
 647         bp->hw.state = PERF_HES_STOPPED;
 648 }
 649
 650 static struct pmu perf_breakpoint = {
 651         .task_ctx_nr    = perf_sw_context, /* could eventually get its own */
 652
 653         .event_init     = hw_breakpoint_event_init,
 654         .add            = hw_breakpoint_add,
 655         .del            = hw_breakpoint_del,
 656         .start          = hw_breakpoint_start,
 657         .stop           = hw_breakpoint_stop,
 658         .read           = hw_breakpoint_pmu_read,
 659 };
 660
 661 int __init init_hw_breakpoint(void)
 662 {
 663         int cpu, err_cpu;
 664         int i;
 665
 666         for (i = 0; i < TYPE_MAX; i++)
 667                 nr_slots[i] = hw_breakpoint_slots(i);
 668
 669         for_each_possible_cpu(cpu) {
 670                 for (i = 0; i < TYPE_MAX; i++) {
 671                         struct bp_cpuinfo *info = get_bp_info(cpu, i);
 672
 673                         info->tsk_pinned = kcalloc(nr_slots[i], sizeof(int),
 674                                                         GFP_KERNEL);
 675                         if (!info->tsk_pinned)
 676                                 goto err_alloc;
 677                 }
 678         }
 679
 680         constraints_initialized = 1;
 681
 682         perf_pmu_register(&perf_breakpoint, "breakpoint", PERF_TYPE_BREAKPOINT);
 683
 684         return register_die_notifier(&hw_breakpoint_exceptions_nb);
 685
 686  err_alloc:
 687         for_each_possible_cpu(err_cpu) {
 688                 for (i = 0; i < TYPE_MAX; i++)
 689                         kfree(get_bp_info(err_cpu, i)->tsk_pinned);
 690                 if (err_cpu == cpu)
 691                         break;
 692         }
 693
 694         return -ENOMEM;
 695 }
 696
 697