arch/arm/kernel/kprobes.c

   1 /*
   2  * arch/arm/kernel/kprobes.c
   3  *
   4  * Kprobes on ARM
   5  *
   6  * Abhishek Sagar <sagar.abhishek@gmail.com>
   7  * Copyright (C) 2006, 2007 Motorola Inc.
   8  *
   9  * Nicolas Pitre <nico@marvell.com>
  10  * Copyright (C) 2007 Marvell Ltd.
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU General Public License version 2 as
  14  * published by the Free Software Foundation.
  15  *
  16  * This program is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * General Public License for more details.
  20  */
  21
  22 #include <linux/kernel.h>
  23 #include <linux/kprobes.h>
  24 #include <linux/module.h>
  25 #include <linux/slab.h>
  26 #include <linux/stop_machine.h>
  27 #include <linux/stringify.h>
  28 #include <asm/traps.h>
  29 #include <asm/cacheflush.h>
  30
  31 #include "kprobes.h"
  32
  33 #define MIN_STACK_SIZE(addr)                            \
  34         min((unsigned long)MAX_STACK_SIZE,              \
  35             (unsigned long)current_thread_info() + THREAD_START_SP - (addr))
  36
  37 #define flush_insns(addr, cnt)                          \
  38         flush_icache_range((unsigned long)(addr),       \
  39                            (unsigned long)(addr) +      \
  40                            sizeof(kprobe_opcode_t) * (cnt))
  41
  42 /* Used as a marker in ARM_pc to note when we're in a jprobe. */
  43 #define JPROBE_MAGIC_ADDR               0xffffffff
  44
  45 DEFINE_PER_CPU(struct kprobe *, current_kprobe) = NULL;
  46 DEFINE_PER_CPU(struct kprobe_ctlblk, kprobe_ctlblk);
  47
  48
  49 int __kprobes arch_prepare_kprobe(struct kprobe *p)
  50 {
  51         kprobe_opcode_t insn;
  52         kprobe_opcode_t tmp_insn[MAX_INSN_SIZE];
  53         unsigned long addr = (unsigned long)p->addr;
  54         int is;
  55
  56         if (addr & 0x3 || in_exception_text(addr))
  57                 return -EINVAL;
  58
  59         insn = *p->addr;
  60         p->opcode = insn;
  61         p->ainsn.insn = tmp_insn;
  62
  63         switch (arm_kprobe_decode_insn(insn, &p->ainsn)) {
  64         case INSN_REJECTED:     /* not supported */
  65                 return -EINVAL;
  66
  67         case INSN_GOOD:         /* instruction uses slot */
  68                 p->ainsn.insn = get_insn_slot();
  69                 if (!p->ainsn.insn)
  70                         return -ENOMEM;
  71                 for (is = 0; is < MAX_INSN_SIZE; ++is)
  72                         p->ainsn.insn[is] = tmp_insn[is];
  73                 flush_insns(p->ainsn.insn, MAX_INSN_SIZE);
  74                 break;
  75
  76         case INSN_GOOD_NO_SLOT: /* instruction doesn't need insn slot */
  77                 p->ainsn.insn = NULL;
  78                 break;
  79         }
  80
  81         return 0;
  82 }
  83
  84 void __kprobes arch_arm_kprobe(struct kprobe *p)
  85 {
  86         *p->addr = KPROBE_BREAKPOINT_INSTRUCTION;
  87         flush_insns(p->addr, 1);
  88 }
  89
  90 /*
  91  * The actual disarming is done here on each CPU and synchronized using
  92  * stop_machine. This synchronization is necessary on SMP to avoid removing
  93  * a probe between the moment the 'Undefined Instruction' exception is raised
  94  * and the moment the exception handler reads the faulting instruction from
  95  * memory.
  96  */
  97 int __kprobes __arch_disarm_kprobe(void *p)
  98 {
  99         struct kprobe *kp = p;
 100         *kp->addr = kp->opcode;
 101         flush_insns(kp->addr, 1);
 102         return 0;
 103 }
 104
 105 void __kprobes arch_disarm_kprobe(struct kprobe *p)
 106 {
 107         stop_machine(__arch_disarm_kprobe, p, &cpu_online_map);
 108 }
 109
 110 void __kprobes arch_remove_kprobe(struct kprobe *p)
 111 {
 112         if (p->ainsn.insn) {
 113                 free_insn_slot(p->ainsn.insn, 0);
 114                 p->ainsn.insn = NULL;
 115         }
 116 }
 117
 118 static void __kprobes save_previous_kprobe(struct kprobe_ctlblk *kcb)
 119 {
 120         kcb->prev_kprobe.kp = kprobe_running();
 121         kcb->prev_kprobe.status = kcb->kprobe_status;
 122 }
 123
 124 static void __kprobes restore_previous_kprobe(struct kprobe_ctlblk *kcb)
 125 {
 126         __get_cpu_var(current_kprobe) = kcb->prev_kprobe.kp;
 127         kcb->kprobe_status = kcb->prev_kprobe.status;
 128 }
 129
 130 static void __kprobes set_current_kprobe(struct kprobe *p)
 131 {
 132         __get_cpu_var(current_kprobe) = p;
 133 }
 134
 135 static void __kprobes singlestep(struct kprobe *p, struct pt_regs *regs,
 136                                  struct kprobe_ctlblk *kcb)
 137 {
 138         regs->ARM_pc += 4;
 139         if (p->ainsn.insn_check_cc(regs->ARM_cpsr))
 140                 p->ainsn.insn_handler(p, regs);
 141 }
 142
 143 /*
 144  * Called with IRQs disabled. IRQs must remain disabled from that point
 145  * all the way until processing this kprobe is complete.  The current
 146  * kprobes implementation cannot process more than one nested level of
 147  * kprobe, and that level is reserved for user kprobe handlers, so we can't
 148  * risk encountering a new kprobe in an interrupt handler.
 149  */
 150 void __kprobes kprobe_handler(struct pt_regs *regs)
 151 {
 152         struct kprobe *p, *cur;
 153         struct kprobe_ctlblk *kcb;
 154         kprobe_opcode_t *addr = (kprobe_opcode_t *)regs->ARM_pc;
 155
 156         kcb = get_kprobe_ctlblk();
 157         cur = kprobe_running();
 158         p = get_kprobe(addr);
 159
 160         if (p) {
 161                 if (cur) {
 162                         /* Kprobe is pending, so we're recursing. */
 163                         switch (kcb->kprobe_status) {
 164                         case KPROBE_HIT_ACTIVE:
 165                         case KPROBE_HIT_SSDONE:
 166                                 /* A pre- or post-handler probe got us here. */
 167                                 kprobes_inc_nmissed_count(p);
 168                                 save_previous_kprobe(kcb);
 169                                 set_current_kprobe(p);
 170                                 kcb->kprobe_status = KPROBE_REENTER;
 171                                 singlestep(p, regs, kcb);
 172                                 restore_previous_kprobe(kcb);
 173                                 break;
 174                         default:
 175                                 /* impossible cases */
 176                                 BUG();
 177                         }
 178                 } else {
 179                         set_current_kprobe(p);
 180                         kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 181
 182                         /*
 183                          * If we have no pre-handler or it returned 0, we
 184                          * continue with normal processing.  If we have a
 185                          * pre-handler and it returned non-zero, it prepped
 186                          * for calling the break_handler below on re-entry,
 187                          * so get out doing nothing more here.
 188                          */
 189                         if (!p->pre_handler || !p->pre_handler(p, regs)) {
 190                                 kcb->kprobe_status = KPROBE_HIT_SS;
 191                                 singlestep(p, regs, kcb);
 192                                 if (p->post_handler) {
 193                                         kcb->kprobe_status = KPROBE_HIT_SSDONE;
 194                                         p->post_handler(p, regs, 0);
 195                                 }
 196                                 reset_current_kprobe();
 197                         }
 198                 }
 199         } else if (cur) {
 200                 /* We probably hit a jprobe.  Call its break handler. */
 201                 if (cur->break_handler && cur->break_handler(cur, regs)) {
 202                         kcb->kprobe_status = KPROBE_HIT_SS;
 203                         singlestep(cur, regs, kcb);
 204                         if (cur->post_handler) {
 205                                 kcb->kprobe_status = KPROBE_HIT_SSDONE;
 206                                 cur->post_handler(cur, regs, 0);
 207                         }
 208                 }
 209                 reset_current_kprobe();
 210         } else {
 211                 /*
 212                  * The probe was removed and a race is in progress.
 213                  * There is nothing we can do about it.  Let's restart
 214                  * the instruction.  By the time we can restart, the
 215                  * real instruction will be there.
 216                  */
 217         }
 218 }
 219
 220 static int __kprobes kprobe_trap_handler(struct pt_regs *regs, unsigned int instr)
 221 {
 222         unsigned long flags;
 223         local_irq_save(flags);
 224         kprobe_handler(regs);
 225         local_irq_restore(flags);
 226         return 0;
 227 }
 228
 229 int __kprobes kprobe_fault_handler(struct pt_regs *regs, unsigned int fsr)
 230 {
 231         struct kprobe *cur = kprobe_running();
 232         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 233
 234         switch (kcb->kprobe_status) {
 235         case KPROBE_HIT_SS:
 236         case KPROBE_REENTER:
 237                 /*
 238                  * We are here because the instruction being single
 239                  * stepped caused a page fault. We reset the current
 240                  * kprobe and the PC to point back to the probe address
 241                  * and allow the page fault handler to continue as a
 242                  * normal page fault.
 243                  */
 244                 regs->ARM_pc = (long)cur->addr;
 245                 if (kcb->kprobe_status == KPROBE_REENTER) {
 246                         restore_previous_kprobe(kcb);
 247                 } else {
 248                         reset_current_kprobe();
 249                 }
 250                 break;
 251
 252         case KPROBE_HIT_ACTIVE:
 253         case KPROBE_HIT_SSDONE:
 254                 /*
 255                  * We increment the nmissed count for accounting,
 256                  * we can also use npre/npostfault count for accounting
 257                  * these specific fault cases.
 258                  */
 259                 kprobes_inc_nmissed_count(cur);
 260
 261                 /*
 262                  * We come here because instructions in the pre/post
 263                  * handler caused the page_fault, this could happen
 264                  * if handler tries to access user space by
 265                  * copy_from_user(), get_user() etc. Let the
 266                  * user-specified handler try to fix it.
 267                  */
 268                 if (cur->fault_handler && cur->fault_handler(cur, regs, fsr))
 269                         return 1;
 270                 break;
 271
 272         default:
 273                 break;
 274         }
 275
 276         return 0;
 277 }
 278
 279 int __kprobes kprobe_exceptions_notify(struct notifier_block *self,
 280                                        unsigned long val, void *data)
 281 {
 282         /*
 283          * notify_die() is currently never called on ARM,
 284          * so this callback is currently empty.
 285          */
 286         return NOTIFY_DONE;
 287 }
 288
 289 /*
 290  * When a retprobed function returns, trampoline_handler() is called,
 291  * calling the kretprobe's handler. We construct a struct pt_regs to
 292  * give a view of registers r0-r11 to the user return-handler.  This is
 293  * not a complete pt_regs structure, but that should be plenty sufficient
 294  * for kretprobe handlers which should normally be interested in r0 only
 295  * anyway.
 296  */
 297 void __naked __kprobes kretprobe_trampoline(void)
 298 {
 299         __asm__ __volatile__ (
 300                 "stmdb  sp!, {r0 - r11}         \n\t"
 301                 "mov    r0, sp                  \n\t"
 302                 "bl     trampoline_handler      \n\t"
 303                 "mov    lr, r0                  \n\t"
 304                 "ldmia  sp!, {r0 - r11}         \n\t"
 305                 "mov    pc, lr                  \n\t"
 306                 : : : "memory");
 307 }
 308
 309 /* Called from kretprobe_trampoline */
 310 static __used __kprobes void *trampoline_handler(struct pt_regs *regs)
 311 {
 312         struct kretprobe_instance *ri = NULL;
 313         struct hlist_head *head, empty_rp;
 314         struct hlist_node *node, *tmp;
 315         unsigned long flags, orig_ret_address = 0;
 316         unsigned long trampoline_address = (unsigned long)&kretprobe_trampoline;
 317
 318         INIT_HLIST_HEAD(&empty_rp);
 319         kretprobe_hash_lock(current, &head, &flags);
 320
 321         /*
 322          * It is possible to have multiple instances associated with a given
 323          * task either because multiple functions in the call path have
 324          * a return probe installed on them, and/or more than one return
 325          * probe was registered for a target function.
 326          *
 327          * We can handle this because:
 328          *     - instances are always inserted at the head of the list
 329          *     - when multiple return probes are registered for the same
 330          *       function, the first instance's ret_addr will point to the
 331          *       real return address, and all the rest will point to
 332          *       kretprobe_trampoline
 333          */
 334         hlist_for_each_entry_safe(ri, node, tmp, head, hlist) {
 335                 if (ri->task != current)
 336                         /* another task is sharing our hash bucket */
 337                         continue;
 338
 339                 if (ri->rp && ri->rp->handler) {
 340                         __get_cpu_var(current_kprobe) = &ri->rp->kp;
 341                         get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 342                         ri->rp->handler(ri, regs);
 343                         __get_cpu_var(current_kprobe) = NULL;
 344                 }
 345
 346                 orig_ret_address = (unsigned long)ri->ret_addr;
 347                 recycle_rp_inst(ri, &empty_rp);
 348
 349                 if (orig_ret_address != trampoline_address)
 350                         /*
 351                          * This is the real return address. Any other
 352                          * instances associated with this task are for
 353                          * other calls deeper on the call stack
 354                          */
 355                         break;
 356         }
 357
 358         kretprobe_assert(ri, orig_ret_address, trampoline_address);
 359         kretprobe_hash_unlock(current, &flags);
 360
 361         hlist_for_each_entry_safe(ri, node, tmp, &empty_rp, hlist) {
 362                 hlist_del(&ri->hlist);
 363                 kfree(ri);
 364         }
 365
 366         return (void *)orig_ret_address;
 367 }
 368
 369 void __kprobes arch_prepare_kretprobe(struct kretprobe_instance *ri,
 370                                       struct pt_regs *regs)
 371 {
 372         ri->ret_addr = (kprobe_opcode_t *)regs->ARM_lr;
 373
 374         /* Replace the return addr with trampoline addr. */
 375         regs->ARM_lr = (unsigned long)&kretprobe_trampoline;
 376 }
 377
 378 int __kprobes setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 379 {
 380         struct jprobe *jp = container_of(p, struct jprobe, kp);
 381         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 382         long sp_addr = regs->ARM_sp;
 383
 384         kcb->jprobe_saved_regs = *regs;
 385         memcpy(kcb->jprobes_stack, (void *)sp_addr, MIN_STACK_SIZE(sp_addr));
 386         regs->ARM_pc = (long)jp->entry;
 387         regs->ARM_cpsr |= PSR_I_BIT;
 388         preempt_disable();
 389         return 1;
 390 }
 391
 392 void __kprobes jprobe_return(void)
 393 {
 394         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 395
 396         __asm__ __volatile__ (
 397                 /*
 398                  * Setup an empty pt_regs. Fill SP and PC fields as
 399                  * they're needed by longjmp_break_handler.
 400                  *
 401                  * We allocate some slack between the original SP and start of
 402                  * our fabricated regs. To be precise we want to have worst case
 403                  * covered which is STMFD with all 16 regs so we allocate 2 *
 404                  * sizeof(struct_pt_regs)).
 405                  *
 406                  * This is to prevent any simulated instruction from writing
 407                  * over the regs when they are accessing the stack.
 408                  */
 409                 "sub    sp, %0, %1              \n\t"
 410                 "ldr    r0, ="__stringify(JPROBE_MAGIC_ADDR)"\n\t"
 411                 "str    %0, [sp, %2]            \n\t"
 412                 "str    r0, [sp, %3]            \n\t"
 413                 "mov    r0, sp                  \n\t"
 414                 "bl     kprobe_handler          \n\t"
 415
 416                 /*
 417                  * Return to the context saved by setjmp_pre_handler
 418                  * and restored by longjmp_break_handler.
 419                  */
 420                 "ldr    r0, [sp, %4]            \n\t"
 421                 "msr    cpsr_cxsf, r0           \n\t"
 422                 "ldmia  sp, {r0 - pc}           \n\t"
 423                 :
 424                 : "r" (kcb->jprobe_saved_regs.ARM_sp),
 425                   "I" (sizeof(struct pt_regs) * 2),
 426                   "J" (offsetof(struct pt_regs, ARM_sp)),
 427                   "J" (offsetof(struct pt_regs, ARM_pc)),
 428                   "J" (offsetof(struct pt_regs, ARM_cpsr))
 429                 : "memory", "cc");
 430 }
 431
 432 int __kprobes longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 433 {
 434         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 435         long stack_addr = kcb->jprobe_saved_regs.ARM_sp;
 436         long orig_sp = regs->ARM_sp;
 437         struct jprobe *jp = container_of(p, struct jprobe, kp);
 438
 439         if (regs->ARM_pc == JPROBE_MAGIC_ADDR) {
 440                 if (orig_sp != stack_addr) {
 441                         struct pt_regs *saved_regs =
 442                                 (struct pt_regs *)kcb->jprobe_saved_regs.ARM_sp;
 443                         printk("current sp %lx does not match saved sp %lx\n",
 444                                orig_sp, stack_addr);
 445                         printk("Saved registers for jprobe %p\n", jp);
 446                         show_regs(saved_regs);
 447                         printk("Current registers\n");
 448                         show_regs(regs);
 449                         BUG();
 450                 }
 451                 *regs = kcb->jprobe_saved_regs;
 452                 memcpy((void *)stack_addr, kcb->jprobes_stack,
 453                        MIN_STACK_SIZE(stack_addr));
 454                 preempt_enable_no_resched();
 455                 return 1;
 456         }
 457         return 0;
 458 }
 459
 460 int __kprobes arch_trampoline_kprobe(struct kprobe *p)
 461 {
 462         return 0;
 463 }
 464
 465 static struct undef_hook kprobes_break_hook = {
 466         .instr_mask     = 0xffffffff,
 467         .instr_val      = KPROBE_BREAKPOINT_INSTRUCTION,
 468         .cpsr_mask      = MODE_MASK,
 469         .cpsr_val       = SVC_MODE,
 470         .fn             = kprobe_trap_handler,
 471 };
 472
 473 int __init arch_init_kprobes()
 474 {
 475         arm_kprobe_decode_init();
 476         register_undef_hook(&kprobes_break_hook);
 477         return 0;
 478 }