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