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/smp.h>
  44
  45 #include <linux/hw_breakpoint.h>
  46
  47 /*
  48  * Constraints data
  49  */
  50
  51 /* Number of pinned cpu breakpoints in a cpu */
  52 static DEFINE_PER_CPU(unsigned int, nr_cpu_bp_pinned);
  53
  54 /* Number of pinned task breakpoints in a cpu */
  55 static DEFINE_PER_CPU(unsigned int, nr_task_bp_pinned[HBP_NUM]);
  56
  57 /* Number of non-pinned cpu/task breakpoints in a cpu */
  58 static DEFINE_PER_CPU(unsigned int, nr_bp_flexible);
  59
  60 /* Gather the number of total pinned and un-pinned bp in a cpuset */
  61 struct bp_busy_slots {
  62         unsigned int pinned;
  63         unsigned int flexible;
  64 };
  65
  66 /* Serialize accesses to the above constraints */
  67 static DEFINE_MUTEX(nr_bp_mutex);
  68
  69 /*
  70  * Report the maximum number of pinned breakpoints a task
  71  * have in this cpu
  72  */
  73 static unsigned int max_task_bp_pinned(int cpu)
  74 {
  75         int i;
  76         unsigned int *tsk_pinned = per_cpu(nr_task_bp_pinned, cpu);
  77
  78         for (i = HBP_NUM -1; i >= 0; i--) {
  79                 if (tsk_pinned[i] > 0)
  80                         return i + 1;
  81         }
  82
  83         return 0;
  84 }
  85
  86 static int task_bp_pinned(struct task_struct *tsk)
  87 {
  88         struct perf_event_context *ctx = tsk->perf_event_ctxp;
  89         struct list_head *list;
  90         struct perf_event *bp;
  91         unsigned long flags;
  92         int count = 0;
  93
  94         if (WARN_ONCE(!ctx, "No perf context for this task"))
  95                 return 0;
  96
  97         list = &ctx->event_list;
  98
  99         raw_spin_lock_irqsave(&ctx->lock, flags);
 100
 101         /*
 102          * The current breakpoint counter is not included in the list
 103          * at the open() callback time
 104          */
 105         list_for_each_entry(bp, list, event_entry) {
 106                 if (bp->attr.type == PERF_TYPE_BREAKPOINT)
 107                         count++;
 108         }
 109
 110         raw_spin_unlock_irqrestore(&ctx->lock, flags);
 111
 112         return count;
 113 }
 114
 115 /*
 116  * Report the number of pinned/un-pinned breakpoints we have in
 117  * a given cpu (cpu > -1) or in all of them (cpu = -1).
 118  */
 119 static void
 120 fetch_bp_busy_slots(struct bp_busy_slots *slots, struct perf_event *bp)
 121 {
 122         int cpu = bp->cpu;
 123         struct task_struct *tsk = bp->ctx->task;
 124
 125         if (cpu >= 0) {
 126                 slots->pinned = per_cpu(nr_cpu_bp_pinned, cpu);
 127                 if (!tsk)
 128                         slots->pinned += max_task_bp_pinned(cpu);
 129                 else
 130                         slots->pinned += task_bp_pinned(tsk);
 131                 slots->flexible = per_cpu(nr_bp_flexible, cpu);
 132
 133                 return;
 134         }
 135
 136         for_each_online_cpu(cpu) {
 137                 unsigned int nr;
 138
 139                 nr = per_cpu(nr_cpu_bp_pinned, cpu);
 140                 if (!tsk)
 141                         nr += max_task_bp_pinned(cpu);
 142                 else
 143                         nr += task_bp_pinned(tsk);
 144
 145                 if (nr > slots->pinned)
 146                         slots->pinned = nr;
 147
 148                 nr = per_cpu(nr_bp_flexible, cpu);
 149
 150                 if (nr > slots->flexible)
 151                         slots->flexible = nr;
 152         }
 153 }
 154
 155 /*
 156  * Add a pinned breakpoint for the given task in our constraint table
 157  */
 158 static void toggle_bp_task_slot(struct task_struct *tsk, int cpu, bool enable)
 159 {
 160         unsigned int *tsk_pinned;
 161         int count = 0;
 162
 163         count = task_bp_pinned(tsk);
 164
 165         tsk_pinned = per_cpu(nr_task_bp_pinned, cpu);
 166         if (enable) {
 167                 tsk_pinned[count]++;
 168                 if (count > 0)
 169                         tsk_pinned[count-1]--;
 170         } else {
 171                 tsk_pinned[count]--;
 172                 if (count > 0)
 173                         tsk_pinned[count-1]++;
 174         }
 175 }
 176
 177 /*
 178  * Add/remove the given breakpoint in our constraint table
 179  */
 180 static void toggle_bp_slot(struct perf_event *bp, bool enable)
 181 {
 182         int cpu = bp->cpu;
 183         struct task_struct *tsk = bp->ctx->task;
 184
 185         /* Pinned counter task profiling */
 186         if (tsk) {
 187                 if (cpu >= 0) {
 188                         toggle_bp_task_slot(tsk, cpu, enable);
 189                         return;
 190                 }
 191
 192                 for_each_online_cpu(cpu)
 193                         toggle_bp_task_slot(tsk, cpu, enable);
 194                 return;
 195         }
 196
 197         /* Pinned counter cpu profiling */
 198         if (enable)
 199                 per_cpu(nr_cpu_bp_pinned, bp->cpu)++;
 200         else
 201                 per_cpu(nr_cpu_bp_pinned, bp->cpu)--;
 202 }
 203
 204 /*
 205  * Contraints to check before allowing this new breakpoint counter:
 206  *
 207  *  == Non-pinned counter == (Considered as pinned for now)
 208  *
 209  *   - If attached to a single cpu, check:
 210  *
 211  *       (per_cpu(nr_bp_flexible, cpu) || (per_cpu(nr_cpu_bp_pinned, cpu)
 212  *           + max(per_cpu(nr_task_bp_pinned, cpu)))) < HBP_NUM
 213  *
 214  *       -> If there are already non-pinned counters in this cpu, it means
 215  *          there is already a free slot for them.
 216  *          Otherwise, we check that the maximum number of per task
 217  *          breakpoints (for this cpu) plus the number of per cpu breakpoint
 218  *          (for this cpu) doesn't cover every registers.
 219  *
 220  *   - If attached to every cpus, check:
 221  *
 222  *       (per_cpu(nr_bp_flexible, *) || (max(per_cpu(nr_cpu_bp_pinned, *))
 223  *           + max(per_cpu(nr_task_bp_pinned, *)))) < HBP_NUM
 224  *
 225  *       -> This is roughly the same, except we check the number of per cpu
 226  *          bp for every cpu and we keep the max one. Same for the per tasks
 227  *          breakpoints.
 228  *
 229  *
 230  * == Pinned counter ==
 231  *
 232  *   - If attached to a single cpu, check:
 233  *
 234  *       ((per_cpu(nr_bp_flexible, cpu) > 1) + per_cpu(nr_cpu_bp_pinned, cpu)
 235  *            + max(per_cpu(nr_task_bp_pinned, cpu))) < HBP_NUM
 236  *
 237  *       -> Same checks as before. But now the nr_bp_flexible, if any, must keep
 238  *          one register at least (or they will never be fed).
 239  *
 240  *   - If attached to every cpus, check:
 241  *
 242  *       ((per_cpu(nr_bp_flexible, *) > 1) + max(per_cpu(nr_cpu_bp_pinned, *))
 243  *            + max(per_cpu(nr_task_bp_pinned, *))) < HBP_NUM
 244  */
 245 int reserve_bp_slot(struct perf_event *bp)
 246 {
 247         struct bp_busy_slots slots = {0};
 248         int ret = 0;
 249
 250         mutex_lock(&nr_bp_mutex);
 251
 252         fetch_bp_busy_slots(&slots, bp);
 253
 254         /* Flexible counters need to keep at least one slot */
 255         if (slots.pinned + (!!slots.flexible) == HBP_NUM) {
 256                 ret = -ENOSPC;
 257                 goto end;
 258         }
 259
 260         toggle_bp_slot(bp, true);
 261
 262 end:
 263         mutex_unlock(&nr_bp_mutex);
 264
 265         return ret;
 266 }
 267
 268 void release_bp_slot(struct perf_event *bp)
 269 {
 270         mutex_lock(&nr_bp_mutex);
 271
 272         toggle_bp_slot(bp, false);
 273
 274         mutex_unlock(&nr_bp_mutex);
 275 }
 276
 277
 278 int register_perf_hw_breakpoint(struct perf_event *bp)
 279 {
 280         int ret;
 281
 282         ret = reserve_bp_slot(bp);
 283         if (ret)
 284                 return ret;
 285
 286         /*
 287          * Ptrace breakpoints can be temporary perf events only
 288          * meant to reserve a slot. In this case, it is created disabled and
 289          * we don't want to check the params right now (as we put a null addr)
 290          * But perf tools create events as disabled and we want to check
 291          * the params for them.
 292          * This is a quick hack that will be removed soon, once we remove
 293          * the tmp breakpoints from ptrace
 294          */
 295         if (!bp->attr.disabled || !bp->overflow_handler)
 296                 ret = arch_validate_hwbkpt_settings(bp, bp->ctx->task);
 297
 298         return ret;
 299 }
 300
 301 /**
 302  * register_user_hw_breakpoint - register a hardware breakpoint for user space
 303  * @attr: breakpoint attributes
 304  * @triggered: callback to trigger when we hit the breakpoint
 305  * @tsk: pointer to 'task_struct' of the process to which the address belongs
 306  */
 307 struct perf_event *
 308 register_user_hw_breakpoint(struct perf_event_attr *attr,
 309                             perf_overflow_handler_t triggered,
 310                             struct task_struct *tsk)
 311 {
 312         return perf_event_create_kernel_counter(attr, -1, tsk->pid, triggered);
 313 }
 314 EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
 315
 316 /**
 317  * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
 318  * @bp: the breakpoint structure to modify
 319  * @attr: new breakpoint attributes
 320  * @triggered: callback to trigger when we hit the breakpoint
 321  * @tsk: pointer to 'task_struct' of the process to which the address belongs
 322  */
 323 int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
 324 {
 325         u64 old_addr = bp->attr.bp_addr;
 326         int old_type = bp->attr.bp_type;
 327         int old_len = bp->attr.bp_len;
 328         int err = 0;
 329
 330         perf_event_disable(bp);
 331
 332         bp->attr.bp_addr = attr->bp_addr;
 333         bp->attr.bp_type = attr->bp_type;
 334         bp->attr.bp_len = attr->bp_len;
 335
 336         if (attr->disabled)
 337                 goto end;
 338
 339         err = arch_validate_hwbkpt_settings(bp, bp->ctx->task);
 340         if (!err)
 341                 perf_event_enable(bp);
 342
 343         if (err) {
 344                 bp->attr.bp_addr = old_addr;
 345                 bp->attr.bp_type = old_type;
 346                 bp->attr.bp_len = old_len;
 347                 if (!bp->attr.disabled)
 348                         perf_event_enable(bp);
 349
 350                 return err;
 351         }
 352
 353 end:
 354         bp->attr.disabled = attr->disabled;
 355
 356         return 0;
 357 }
 358 EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
 359
 360 /**
 361  * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
 362  * @bp: the breakpoint structure to unregister
 363  */
 364 void unregister_hw_breakpoint(struct perf_event *bp)
 365 {
 366         if (!bp)
 367                 return;
 368         perf_event_release_kernel(bp);
 369 }
 370 EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
 371
 372 /**
 373  * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
 374  * @attr: breakpoint attributes
 375  * @triggered: callback to trigger when we hit the breakpoint
 376  *
 377  * @return a set of per_cpu pointers to perf events
 378  */
 379 struct perf_event **
 380 register_wide_hw_breakpoint(struct perf_event_attr *attr,
 381                             perf_overflow_handler_t triggered)
 382 {
 383         struct perf_event **cpu_events, **pevent, *bp;
 384         long err;
 385         int cpu;
 386
 387         cpu_events = alloc_percpu(typeof(*cpu_events));
 388         if (!cpu_events)
 389                 return ERR_PTR(-ENOMEM);
 390
 391         for_each_possible_cpu(cpu) {
 392                 pevent = per_cpu_ptr(cpu_events, cpu);
 393                 bp = perf_event_create_kernel_counter(attr, cpu, -1, triggered);
 394
 395                 *pevent = bp;
 396
 397                 if (IS_ERR(bp)) {
 398                         err = PTR_ERR(bp);
 399                         goto fail;
 400                 }
 401         }
 402
 403         return cpu_events;
 404
 405 fail:
 406         for_each_possible_cpu(cpu) {
 407                 pevent = per_cpu_ptr(cpu_events, cpu);
 408                 if (IS_ERR(*pevent))
 409                         break;
 410                 unregister_hw_breakpoint(*pevent);
 411         }
 412         free_percpu(cpu_events);
 413         /* return the error if any */
 414         return ERR_PTR(err);
 415 }
 416 EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
 417
 418 /**
 419  * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
 420  * @cpu_events: the per cpu set of events to unregister
 421  */
 422 void unregister_wide_hw_breakpoint(struct perf_event **cpu_events)
 423 {
 424         int cpu;
 425         struct perf_event **pevent;
 426
 427         for_each_possible_cpu(cpu) {
 428                 pevent = per_cpu_ptr(cpu_events, cpu);
 429                 unregister_hw_breakpoint(*pevent);
 430         }
 431         free_percpu(cpu_events);
 432 }
 433 EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
 434
 435 static struct notifier_block hw_breakpoint_exceptions_nb = {
 436         .notifier_call = hw_breakpoint_exceptions_notify,
 437         /* we need to be notified first */
 438         .priority = 0x7fffffff
 439 };
 440
 441 static int __init init_hw_breakpoint(void)
 442 {
 443         return register_die_notifier(&hw_breakpoint_exceptions_nb);
 444 }
 445 core_initcall(init_hw_breakpoint);
 446
 447
 448 struct pmu perf_ops_bp = {
 449         .enable         = arch_install_hw_breakpoint,
 450         .disable        = arch_uninstall_hw_breakpoint,
 451         .read           = hw_breakpoint_pmu_read,
 452         .unthrottle     = hw_breakpoint_pmu_unthrottle
 453 };