kernel/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(struct perf_event *bp, enum bp_type_idx type)
 115 {
 116         struct perf_event_context *ctx = bp->ctx;
 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->ctx == ctx && find_slot_idx(iter) == type)
 122                         count += hw_breakpoint_weight(iter);
 123         }
 124
 125         return count;
 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         int cpu = bp->cpu;
 137         struct task_struct *tsk = bp->ctx->task;
 138
 139         if (cpu >= 0) {
 140                 slots->pinned = per_cpu(nr_cpu_bp_pinned[type], cpu);
 141                 if (!tsk)
 142                         slots->pinned += max_task_bp_pinned(cpu, type);
 143                 else
 144                         slots->pinned += task_bp_pinned(bp, type);
 145                 slots->flexible = per_cpu(nr_bp_flexible[type], cpu);
 146
 147                 return;
 148         }
 149
 150         for_each_online_cpu(cpu) {
 151                 unsigned int nr;
 152
 153                 nr = per_cpu(nr_cpu_bp_pinned[type], cpu);
 154                 if (!tsk)
 155                         nr += max_task_bp_pinned(cpu, type);
 156                 else
 157                         nr += task_bp_pinned(bp, type);
 158
 159                 if (nr > slots->pinned)
 160                         slots->pinned = nr;
 161
 162                 nr = per_cpu(nr_bp_flexible[type], cpu);
 163
 164                 if (nr > slots->flexible)
 165                         slots->flexible = nr;
 166         }
 167 }
 168
 169 /*
 170  * For now, continue to consider flexible as pinned, until we can
 171  * ensure no flexible event can ever be scheduled before a pinned event
 172  * in a same cpu.
 173  */
 174 static void
 175 fetch_this_slot(struct bp_busy_slots *slots, int weight)
 176 {
 177         slots->pinned += weight;
 178 }
 179
 180 /*
 181  * Add a pinned breakpoint for the given task in our constraint table
 182  */
 183 static void toggle_bp_task_slot(struct perf_event *bp, int cpu, bool enable,
 184                                 enum bp_type_idx type, int weight)
 185 {
 186         unsigned int *tsk_pinned;
 187         int old_count = 0;
 188         int old_idx = 0;
 189         int idx = 0;
 190
 191         old_count = task_bp_pinned(bp, type);
 192         old_idx = old_count - 1;
 193         idx = old_idx + weight;
 194
 195         /* tsk_pinned[n] is the number of tasks having n breakpoints */
 196         tsk_pinned = per_cpu(nr_task_bp_pinned[type], cpu);
 197         if (enable) {
 198                 tsk_pinned[idx]++;
 199                 if (old_count > 0)
 200                         tsk_pinned[old_idx]--;
 201         } else {
 202                 tsk_pinned[idx]--;
 203                 if (old_count > 0)
 204                         tsk_pinned[old_idx]++;
 205         }
 206 }
 207
 208 /*
 209  * Add/remove the given breakpoint in our constraint table
 210  */
 211 static void
 212 toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type,
 213                int weight)
 214 {
 215         int cpu = bp->cpu;
 216         struct task_struct *tsk = bp->ctx->task;
 217
 218         /* Pinned counter cpu profiling */
 219         if (!tsk) {
 220
 221                 if (enable)
 222                         per_cpu(nr_cpu_bp_pinned[type], bp->cpu) += weight;
 223                 else
 224                         per_cpu(nr_cpu_bp_pinned[type], bp->cpu) -= weight;
 225                 return;
 226         }
 227
 228         /* Pinned counter task profiling */
 229
 230         if (!enable)
 231                 list_del(&bp->hw.bp_list);
 232
 233         if (cpu >= 0) {
 234                 toggle_bp_task_slot(bp, cpu, enable, type, weight);
 235         } else {
 236                 for_each_online_cpu(cpu)
 237                         toggle_bp_task_slot(bp, cpu, enable, type, weight);
 238         }
 239
 240         if (enable)
 241                 list_add_tail(&bp->hw.bp_list, &bp_task_head);
 242 }
 243
 244 /*
 245  * Contraints to check before allowing this new breakpoint counter:
 246  *
 247  *  == Non-pinned counter == (Considered as pinned for now)
 248  *
 249  *   - If attached to a single cpu, check:
 250  *
 251  *       (per_cpu(nr_bp_flexible, cpu) || (per_cpu(nr_cpu_bp_pinned, cpu)
 252  *           + max(per_cpu(nr_task_bp_pinned, cpu)))) < HBP_NUM
 253  *
 254  *       -> If there are already non-pinned counters in this cpu, it means
 255  *          there is already a free slot for them.
 256  *          Otherwise, we check that the maximum number of per task
 257  *          breakpoints (for this cpu) plus the number of per cpu breakpoint
 258  *          (for this cpu) doesn't cover every registers.
 259  *
 260  *   - If attached to every cpus, check:
 261  *
 262  *       (per_cpu(nr_bp_flexible, *) || (max(per_cpu(nr_cpu_bp_pinned, *))
 263  *           + max(per_cpu(nr_task_bp_pinned, *)))) < HBP_NUM
 264  *
 265  *       -> This is roughly the same, except we check the number of per cpu
 266  *          bp for every cpu and we keep the max one. Same for the per tasks
 267  *          breakpoints.
 268  *
 269  *
 270  * == Pinned counter ==
 271  *
 272  *   - If attached to a single cpu, check:
 273  *
 274  *       ((per_cpu(nr_bp_flexible, cpu) > 1) + per_cpu(nr_cpu_bp_pinned, cpu)
 275  *            + max(per_cpu(nr_task_bp_pinned, cpu))) < HBP_NUM
 276  *
 277  *       -> Same checks as before. But now the nr_bp_flexible, if any, must keep
 278  *          one register at least (or they will never be fed).
 279  *
 280  *   - If attached to every cpus, check:
 281  *
 282  *       ((per_cpu(nr_bp_flexible, *) > 1) + max(per_cpu(nr_cpu_bp_pinned, *))
 283  *            + max(per_cpu(nr_task_bp_pinned, *))) < HBP_NUM
 284  */
 285 static int __reserve_bp_slot(struct perf_event *bp)
 286 {
 287         struct bp_busy_slots slots = {0};
 288         enum bp_type_idx type;
 289         int weight;
 290
 291         /* We couldn't initialize breakpoint constraints on boot */
 292         if (!constraints_initialized)
 293                 return -ENOMEM;
 294
 295         /* Basic checks */
 296         if (bp->attr.bp_type == HW_BREAKPOINT_EMPTY ||
 297             bp->attr.bp_type == HW_BREAKPOINT_INVALID)
 298                 return -EINVAL;
 299
 300         type = find_slot_idx(bp);
 301         weight = hw_breakpoint_weight(bp);
 302
 303         fetch_bp_busy_slots(&slots, bp, type);
 304         /*
 305          * Simulate the addition of this breakpoint to the constraints
 306          * and see the result.
 307          */
 308         fetch_this_slot(&slots, weight);
 309
 310         /* Flexible counters need to keep at least one slot */
 311         if (slots.pinned + (!!slots.flexible) > nr_slots[type])
 312                 return -ENOSPC;
 313
 314         toggle_bp_slot(bp, true, type, weight);
 315
 316         return 0;
 317 }
 318
 319 int reserve_bp_slot(struct perf_event *bp)
 320 {
 321         int ret;
 322
 323         mutex_lock(&nr_bp_mutex);
 324
 325         ret = __reserve_bp_slot(bp);
 326
 327         mutex_unlock(&nr_bp_mutex);
 328
 329         return ret;
 330 }
 331
 332 static void __release_bp_slot(struct perf_event *bp)
 333 {
 334         enum bp_type_idx type;
 335         int weight;
 336
 337         type = find_slot_idx(bp);
 338         weight = hw_breakpoint_weight(bp);
 339         toggle_bp_slot(bp, false, type, weight);
 340 }
 341
 342 void release_bp_slot(struct perf_event *bp)
 343 {
 344         mutex_lock(&nr_bp_mutex);
 345
 346         __release_bp_slot(bp);
 347
 348         mutex_unlock(&nr_bp_mutex);
 349 }
 350
 351 /*
 352  * Allow the kernel debugger to reserve breakpoint slots without
 353  * taking a lock using the dbg_* variant of for the reserve and
 354  * release breakpoint slots.
 355  */
 356 int dbg_reserve_bp_slot(struct perf_event *bp)
 357 {
 358         if (mutex_is_locked(&nr_bp_mutex))
 359                 return -1;
 360
 361         return __reserve_bp_slot(bp);
 362 }
 363
 364 int dbg_release_bp_slot(struct perf_event *bp)
 365 {
 366         if (mutex_is_locked(&nr_bp_mutex))
 367                 return -1;
 368
 369         __release_bp_slot(bp);
 370
 371         return 0;
 372 }
 373
 374 static int validate_hw_breakpoint(struct perf_event *bp)
 375 {
 376         int ret;
 377
 378         ret = arch_validate_hwbkpt_settings(bp);
 379         if (ret)
 380                 return ret;
 381
 382         if (arch_check_bp_in_kernelspace(bp)) {
 383                 if (bp->attr.exclude_kernel)
 384                         return -EINVAL;
 385                 /*
 386                  * Don't let unprivileged users set a breakpoint in the trap
 387                  * path to avoid trap recursion attacks.
 388                  */
 389                 if (!capable(CAP_SYS_ADMIN))
 390                         return -EPERM;
 391         }
 392
 393         return 0;
 394 }
 395
 396 int register_perf_hw_breakpoint(struct perf_event *bp)
 397 {
 398         int ret;
 399
 400         ret = reserve_bp_slot(bp);
 401         if (ret)
 402                 return ret;
 403
 404         ret = validate_hw_breakpoint(bp);
 405
 406         /* if arch_validate_hwbkpt_settings() fails then release bp slot */
 407         if (ret)
 408                 release_bp_slot(bp);
 409
 410         return ret;
 411 }
 412
 413 /**
 414  * register_user_hw_breakpoint - register a hardware breakpoint for user space
 415  * @attr: breakpoint attributes
 416  * @triggered: callback to trigger when we hit the breakpoint
 417  * @tsk: pointer to 'task_struct' of the process to which the address belongs
 418  */
 419 struct perf_event *
 420 register_user_hw_breakpoint(struct perf_event_attr *attr,
 421                             perf_overflow_handler_t triggered,
 422                             struct task_struct *tsk)
 423 {
 424         return perf_event_create_kernel_counter(attr, -1, tsk->pid, triggered);
 425 }
 426 EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
 427
 428 /**
 429  * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
 430  * @bp: the breakpoint structure to modify
 431  * @attr: new breakpoint attributes
 432  * @triggered: callback to trigger when we hit the breakpoint
 433  * @tsk: pointer to 'task_struct' of the process to which the address belongs
 434  */
 435 int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
 436 {
 437         u64 old_addr = bp->attr.bp_addr;
 438         u64 old_len = bp->attr.bp_len;
 439         int old_type = bp->attr.bp_type;
 440         int err = 0;
 441
 442         perf_event_disable(bp);
 443
 444         bp->attr.bp_addr = attr->bp_addr;
 445         bp->attr.bp_type = attr->bp_type;
 446         bp->attr.bp_len = attr->bp_len;
 447
 448         if (attr->disabled)
 449                 goto end;
 450
 451         err = validate_hw_breakpoint(bp);
 452         if (!err)
 453                 perf_event_enable(bp);
 454
 455         if (err) {
 456                 bp->attr.bp_addr = old_addr;
 457                 bp->attr.bp_type = old_type;
 458                 bp->attr.bp_len = old_len;
 459                 if (!bp->attr.disabled)
 460                         perf_event_enable(bp);
 461
 462                 return err;
 463         }
 464
 465 end:
 466         bp->attr.disabled = attr->disabled;
 467
 468         return 0;
 469 }
 470 EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
 471
 472 /**
 473  * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
 474  * @bp: the breakpoint structure to unregister
 475  */
 476 void unregister_hw_breakpoint(struct perf_event *bp)
 477 {
 478         if (!bp)
 479                 return;
 480         perf_event_release_kernel(bp);
 481 }
 482 EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
 483
 484 /**
 485  * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
 486  * @attr: breakpoint attributes
 487  * @triggered: callback to trigger when we hit the breakpoint
 488  *
 489  * @return a set of per_cpu pointers to perf events
 490  */
 491 struct perf_event * __percpu *
 492 register_wide_hw_breakpoint(struct perf_event_attr *attr,
 493                             perf_overflow_handler_t triggered)
 494 {
 495         struct perf_event * __percpu *cpu_events, **pevent, *bp;
 496         long err;
 497         int cpu;
 498
 499         cpu_events = alloc_percpu(typeof(*cpu_events));
 500         if (!cpu_events)
 501                 return (void __percpu __force *)ERR_PTR(-ENOMEM);
 502
 503         get_online_cpus();
 504         for_each_online_cpu(cpu) {
 505                 pevent = per_cpu_ptr(cpu_events, cpu);
 506                 bp = perf_event_create_kernel_counter(attr, cpu, -1, triggered);
 507
 508                 *pevent = bp;
 509
 510                 if (IS_ERR(bp)) {
 511                         err = PTR_ERR(bp);
 512                         goto fail;
 513                 }
 514         }
 515         put_online_cpus();
 516
 517         return cpu_events;
 518
 519 fail:
 520         for_each_online_cpu(cpu) {
 521                 pevent = per_cpu_ptr(cpu_events, cpu);
 522                 if (IS_ERR(*pevent))
 523                         break;
 524                 unregister_hw_breakpoint(*pevent);
 525         }
 526         put_online_cpus();
 527
 528         free_percpu(cpu_events);
 529         return (void __percpu __force *)ERR_PTR(err);
 530 }
 531 EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
 532
 533 /**
 534  * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
 535  * @cpu_events: the per cpu set of events to unregister
 536  */
 537 void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
 538 {
 539         int cpu;
 540         struct perf_event **pevent;
 541
 542         for_each_possible_cpu(cpu) {
 543                 pevent = per_cpu_ptr(cpu_events, cpu);
 544                 unregister_hw_breakpoint(*pevent);
 545         }
 546         free_percpu(cpu_events);
 547 }
 548 EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
 549
 550 static struct notifier_block hw_breakpoint_exceptions_nb = {
 551         .notifier_call = hw_breakpoint_exceptions_notify,
 552         /* we need to be notified first */
 553         .priority = 0x7fffffff
 554 };
 555
 556 static int __init init_hw_breakpoint(void)
 557 {
 558         unsigned int **task_bp_pinned;
 559         int cpu, err_cpu;
 560         int i;
 561
 562         for (i = 0; i < TYPE_MAX; i++)
 563                 nr_slots[i] = hw_breakpoint_slots(i);
 564
 565         for_each_possible_cpu(cpu) {
 566                 for (i = 0; i < TYPE_MAX; i++) {
 567                         task_bp_pinned = &per_cpu(nr_task_bp_pinned[i], cpu);
 568                         *task_bp_pinned = kzalloc(sizeof(int) * nr_slots[i],
 569                                                   GFP_KERNEL);
 570                         if (!*task_bp_pinned)
 571                                 goto err_alloc;
 572                 }
 573         }
 574
 575         constraints_initialized = 1;
 576
 577         return register_die_notifier(&hw_breakpoint_exceptions_nb);
 578
 579  err_alloc:
 580         for_each_possible_cpu(err_cpu) {
 581                 if (err_cpu == cpu)
 582                         break;
 583                 for (i = 0; i < TYPE_MAX; i++)
 584                         kfree(per_cpu(nr_task_bp_pinned[i], cpu));
 585         }
 586
 587         return -ENOMEM;
 588 }
 589 core_initcall(init_hw_breakpoint);
 590
 591
 592 struct pmu perf_ops_bp = {
 593         .enable         = arch_install_hw_breakpoint,
 594         .disable        = arch_uninstall_hw_breakpoint,
 595         .read           = hw_breakpoint_pmu_read,
 596 };