git.oblomov.eu Git - linux-2.6/blob - arch/x86_64/kernel/kprobes.c

   1 /*
   2  *  Kernel Probes (KProbes)
   3  *  arch/x86_64/kernel/kprobes.c
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  18  *
  19  * Copyright (C) IBM Corporation, 2002, 2004
  20  *
  21  * 2002-Oct     Created by Vamsi Krishna S <vamsi_krishna@in.ibm.com> Kernel
  22  *              Probes initial implementation ( includes contributions from
  23  *              Rusty Russell).
  24  * 2004-July    Suparna Bhattacharya <suparna@in.ibm.com> added jumper probes
  25  *              interface to access function arguments.
  26  * 2004-Oct     Jim Keniston <kenistoj@us.ibm.com> and Prasanna S Panchamukhi
  27  *              <prasanna@in.ibm.com> adapted for x86_64
  28  * 2005-Mar     Roland McGrath <roland@redhat.com>
  29  *              Fixed to handle %rip-relative addressing mode correctly.
  30  * 2005-May     Rusty Lynch <rusty.lynch@intel.com>
  31  *              Added function return probes functionality
  32  */
  33
  34 #include <linux/config.h>
  35 #include <linux/kprobes.h>
  36 #include <linux/ptrace.h>
  37 #include <linux/spinlock.h>
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include <linux/preempt.h>
  41 #include <linux/moduleloader.h>
  42 #include <asm/cacheflush.h>
  43 #include <asm/pgtable.h>
  44 #include <asm/kdebug.h>
  45
  46 static DECLARE_MUTEX(kprobe_mutex);
  47
  48 static struct kprobe *current_kprobe;
  49 static unsigned long kprobe_status, kprobe_old_rflags, kprobe_saved_rflags;
  50 static struct kprobe *kprobe_prev;
  51 static unsigned long kprobe_status_prev, kprobe_old_rflags_prev, kprobe_saved_rflags_prev;
  52 static struct pt_regs jprobe_saved_regs;
  53 static long *jprobe_saved_rsp;
  54 static kprobe_opcode_t *get_insn_slot(void);
  55 static void free_insn_slot(kprobe_opcode_t *slot);
  56 void jprobe_return_end(void);
  57
  58 /* copy of the kernel stack at the probe fire time */
  59 static kprobe_opcode_t jprobes_stack[MAX_STACK_SIZE];
  60
  61 /*
  62  * returns non-zero if opcode modifies the interrupt flag.
  63  */
  64 static inline int is_IF_modifier(kprobe_opcode_t *insn)
  65 {
  66         switch (*insn) {
  67         case 0xfa:              /* cli */
  68         case 0xfb:              /* sti */
  69         case 0xcf:              /* iret/iretd */
  70         case 0x9d:              /* popf/popfd */
  71                 return 1;
  72         }
  73
  74         if (*insn  >= 0x40 && *insn <= 0x4f && *++insn == 0xcf)
  75                 return 1;
  76         return 0;
  77 }
  78
  79 int arch_prepare_kprobe(struct kprobe *p)
  80 {
  81         /* insn: must be on special executable page on x86_64. */
  82         up(&kprobe_mutex);
  83         p->ainsn.insn = get_insn_slot();
  84         down(&kprobe_mutex);
  85         if (!p->ainsn.insn) {
  86                 return -ENOMEM;
  87         }
  88         return 0;
  89 }
  90
  91 /*
  92  * Determine if the instruction uses the %rip-relative addressing mode.
  93  * If it does, return the address of the 32-bit displacement word.
  94  * If not, return null.
  95  */
  96 static inline s32 *is_riprel(u8 *insn)
  97 {
  98 #define W(row,b0,b1,b2,b3,b4,b5,b6,b7,b8,b9,ba,bb,bc,bd,be,bf)                \
  99         (((b0##UL << 0x0)|(b1##UL << 0x1)|(b2##UL << 0x2)|(b3##UL << 0x3) |   \
 100           (b4##UL << 0x4)|(b5##UL << 0x5)|(b6##UL << 0x6)|(b7##UL << 0x7) |   \
 101           (b8##UL << 0x8)|(b9##UL << 0x9)|(ba##UL << 0xa)|(bb##UL << 0xb) |   \
 102           (bc##UL << 0xc)|(bd##UL << 0xd)|(be##UL << 0xe)|(bf##UL << 0xf))    \
 103          << (row % 64))
 104         static const u64 onebyte_has_modrm[256 / 64] = {
 105                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 106                 /*      -------------------------------         */
 107                 W(0x00, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 00 */
 108                 W(0x10, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 10 */
 109                 W(0x20, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0)| /* 20 */
 110                 W(0x30, 1,1,1,1,0,0,0,0,1,1,1,1,0,0,0,0), /* 30 */
 111                 W(0x40, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 40 */
 112                 W(0x50, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 50 */
 113                 W(0x60, 0,0,1,1,0,0,0,0,0,1,0,1,0,0,0,0)| /* 60 */
 114                 W(0x70, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 70 */
 115                 W(0x80, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 80 */
 116                 W(0x90, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 90 */
 117                 W(0xa0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* a0 */
 118                 W(0xb0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* b0 */
 119                 W(0xc0, 1,1,0,0,1,1,1,1,0,0,0,0,0,0,0,0)| /* c0 */
 120                 W(0xd0, 1,1,1,1,0,0,0,0,1,1,1,1,1,1,1,1)| /* d0 */
 121                 W(0xe0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* e0 */
 122                 W(0xf0, 0,0,0,0,0,0,1,1,0,0,0,0,0,0,1,1)  /* f0 */
 123                 /*      -------------------------------         */
 124                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 125         };
 126         static const u64 twobyte_has_modrm[256 / 64] = {
 127                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 128                 /*      -------------------------------         */
 129                 W(0x00, 1,1,1,1,0,0,0,0,0,0,0,0,0,1,0,1)| /* 0f */
 130                 W(0x10, 1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0)| /* 1f */
 131                 W(0x20, 1,1,1,1,1,0,1,0,1,1,1,1,1,1,1,1)| /* 2f */
 132                 W(0x30, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), /* 3f */
 133                 W(0x40, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 4f */
 134                 W(0x50, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 5f */
 135                 W(0x60, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 6f */
 136                 W(0x70, 1,1,1,1,1,1,1,0,0,0,0,0,1,1,1,1), /* 7f */
 137                 W(0x80, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0)| /* 8f */
 138                 W(0x90, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* 9f */
 139                 W(0xa0, 0,0,0,1,1,1,1,1,0,0,0,1,1,1,1,1)| /* af */
 140                 W(0xb0, 1,1,1,1,1,1,1,1,0,0,1,1,1,1,1,1), /* bf */
 141                 W(0xc0, 1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0)| /* cf */
 142                 W(0xd0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* df */
 143                 W(0xe0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1)| /* ef */
 144                 W(0xf0, 1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,0)  /* ff */
 145                 /*      -------------------------------         */
 146                 /*      0 1 2 3 4 5 6 7 8 9 a b c d e f         */
 147         };
 148 #undef  W
 149         int need_modrm;
 150
 151         /* Skip legacy instruction prefixes.  */
 152         while (1) {
 153                 switch (*insn) {
 154                 case 0x66:
 155                 case 0x67:
 156                 case 0x2e:
 157                 case 0x3e:
 158                 case 0x26:
 159                 case 0x64:
 160                 case 0x65:
 161                 case 0x36:
 162                 case 0xf0:
 163                 case 0xf3:
 164                 case 0xf2:
 165                         ++insn;
 166                         continue;
 167                 }
 168                 break;
 169         }
 170
 171         /* Skip REX instruction prefix.  */
 172         if ((*insn & 0xf0) == 0x40)
 173                 ++insn;
 174
 175         if (*insn == 0x0f) {    /* Two-byte opcode.  */
 176                 ++insn;
 177                 need_modrm = test_bit(*insn, twobyte_has_modrm);
 178         } else {                /* One-byte opcode.  */
 179                 need_modrm = test_bit(*insn, onebyte_has_modrm);
 180         }
 181
 182         if (need_modrm) {
 183                 u8 modrm = *++insn;
 184                 if ((modrm & 0xc7) == 0x05) { /* %rip+disp32 addressing mode */
 185                         /* Displacement follows ModRM byte.  */
 186                         return (s32 *) ++insn;
 187                 }
 188         }
 189
 190         /* No %rip-relative addressing mode here.  */
 191         return NULL;
 192 }
 193
 194 void arch_copy_kprobe(struct kprobe *p)
 195 {
 196         s32 *ripdisp;
 197         memcpy(p->ainsn.insn, p->addr, MAX_INSN_SIZE);
 198         ripdisp = is_riprel(p->ainsn.insn);
 199         if (ripdisp) {
 200                 /*
 201                  * The copied instruction uses the %rip-relative
 202                  * addressing mode.  Adjust the displacement for the
 203                  * difference between the original location of this
 204                  * instruction and the location of the copy that will
 205                  * actually be run.  The tricky bit here is making sure
 206                  * that the sign extension happens correctly in this
 207                  * calculation, since we need a signed 32-bit result to
 208                  * be sign-extended to 64 bits when it's added to the
 209                  * %rip value and yield the same 64-bit result that the
 210                  * sign-extension of the original signed 32-bit
 211                  * displacement would have given.
 212                  */
 213                 s64 disp = (u8 *) p->addr + *ripdisp - (u8 *) p->ainsn.insn;
 214                 BUG_ON((s64) (s32) disp != disp); /* Sanity check.  */
 215                 *ripdisp = disp;
 216         }
 217         p->opcode = *p->addr;
 218 }
 219
 220 void arch_arm_kprobe(struct kprobe *p)
 221 {
 222         *p->addr = BREAKPOINT_INSTRUCTION;
 223         flush_icache_range((unsigned long) p->addr,
 224                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 225 }
 226
 227 void arch_disarm_kprobe(struct kprobe *p)
 228 {
 229         *p->addr = p->opcode;
 230         flush_icache_range((unsigned long) p->addr,
 231                            (unsigned long) p->addr + sizeof(kprobe_opcode_t));
 232 }
 233
 234 void arch_remove_kprobe(struct kprobe *p)
 235 {
 236         up(&kprobe_mutex);
 237         free_insn_slot(p->ainsn.insn);
 238         down(&kprobe_mutex);
 239 }
 240
 241 static inline void save_previous_kprobe(void)
 242 {
 243         kprobe_prev = current_kprobe;
 244         kprobe_status_prev = kprobe_status;
 245         kprobe_old_rflags_prev = kprobe_old_rflags;
 246         kprobe_saved_rflags_prev = kprobe_saved_rflags;
 247 }
 248
 249 static inline void restore_previous_kprobe(void)
 250 {
 251         current_kprobe = kprobe_prev;
 252         kprobe_status = kprobe_status_prev;
 253         kprobe_old_rflags = kprobe_old_rflags_prev;
 254         kprobe_saved_rflags = kprobe_saved_rflags_prev;
 255 }
 256
 257 static inline void set_current_kprobe(struct kprobe *p, struct pt_regs *regs)
 258 {
 259         current_kprobe = p;
 260         kprobe_saved_rflags = kprobe_old_rflags
 261                 = (regs->eflags & (TF_MASK | IF_MASK));
 262         if (is_IF_modifier(p->ainsn.insn))
 263                 kprobe_saved_rflags &= ~IF_MASK;
 264 }
 265
 266 static void prepare_singlestep(struct kprobe *p, struct pt_regs *regs)
 267 {
 268         regs->eflags |= TF_MASK;
 269         regs->eflags &= ~IF_MASK;
 270         /*single step inline if the instruction is an int3*/
 271         if (p->opcode == BREAKPOINT_INSTRUCTION)
 272                 regs->rip = (unsigned long)p->addr;
 273         else
 274                 regs->rip = (unsigned long)p->ainsn.insn;
 275 }
 276
 277 struct task_struct  *arch_get_kprobe_task(void *ptr)
 278 {
 279         return ((struct thread_info *) (((unsigned long) ptr) &
 280                                         (~(THREAD_SIZE -1))))->task;
 281 }
 282
 283 void arch_prepare_kretprobe(struct kretprobe *rp, struct pt_regs *regs)
 284 {
 285         unsigned long *sara = (unsigned long *)regs->rsp;
 286         struct kretprobe_instance *ri;
 287         static void *orig_ret_addr;
 288
 289         /*
 290          * Save the return address when the return probe hits
 291          * the first time, and use it to populate the (krprobe
 292          * instance)->ret_addr for subsequent return probes at
 293          * the same addrress since stack address would have
 294          * the kretprobe_trampoline by then.
 295          */
 296         if (((void*) *sara) != kretprobe_trampoline)
 297                 orig_ret_addr = (void*) *sara;
 298
 299         if ((ri = get_free_rp_inst(rp)) != NULL) {
 300                 ri->rp = rp;
 301                 ri->stack_addr = sara;
 302                 ri->ret_addr = orig_ret_addr;
 303                 add_rp_inst(ri);
 304                 /* Replace the return addr with trampoline addr */
 305                 *sara = (unsigned long) &kretprobe_trampoline;
 306         } else {
 307                 rp->nmissed++;
 308         }
 309 }
 310
 311 void arch_kprobe_flush_task(struct task_struct *tk)
 312 {
 313         struct kretprobe_instance *ri;
 314         while ((ri = get_rp_inst_tsk(tk)) != NULL) {
 315                 *((unsigned long *)(ri->stack_addr)) =
 316                                         (unsigned long) ri->ret_addr;
 317                 recycle_rp_inst(ri);
 318         }
 319 }
 320
 321 /*
 322  * Interrupts are disabled on entry as trap3 is an interrupt gate and they
 323  * remain disabled thorough out this function.
 324  */
 325 int kprobe_handler(struct pt_regs *regs)
 326 {
 327         struct kprobe *p;
 328         int ret = 0;
 329         kprobe_opcode_t *addr = (kprobe_opcode_t *)(regs->rip - sizeof(kprobe_opcode_t));
 330
 331         /* We're in an interrupt, but this is clear and BUG()-safe. */
 332         preempt_disable();
 333
 334         /* Check we're not actually recursing */
 335         if (kprobe_running()) {
 336                 /* We *are* holding lock here, so this is safe.
 337                    Disarm the probe we just hit, and ignore it. */
 338                 p = get_kprobe(addr);
 339                 if (p) {
 340                         if (kprobe_status == KPROBE_HIT_SS) {
 341                                 regs->eflags &= ~TF_MASK;
 342                                 regs->eflags |= kprobe_saved_rflags;
 343                                 unlock_kprobes();
 344                                 goto no_kprobe;
 345                         } else if (kprobe_status == KPROBE_HIT_SSDONE) {
 346                                 /* TODO: Provide re-entrancy from
 347                                  * post_kprobes_handler() and avoid exception
 348                                  * stack corruption while single-stepping on
 349                                  * the instruction of the new probe.
 350                                  */
 351                                 arch_disarm_kprobe(p);
 352                                 regs->rip = (unsigned long)p->addr;
 353                                 ret = 1;
 354                         } else {
 355                                 /* We have reentered the kprobe_handler(), since
 356                                  * another probe was hit while within the
 357                                  * handler. We here save the original kprobe
 358                                  * variables and just single step on instruction
 359                                  * of the new probe without calling any user
 360                                  * handlers.
 361                                  */
 362                                 save_previous_kprobe();
 363                                 set_current_kprobe(p, regs);
 364                                 p->nmissed++;
 365                                 prepare_singlestep(p, regs);
 366                                 kprobe_status = KPROBE_REENTER;
 367                                 return 1;
 368                         }
 369                 } else {
 370                         p = current_kprobe;
 371                         if (p->break_handler && p->break_handler(p, regs)) {
 372                                 goto ss_probe;
 373                         }
 374                 }
 375                 /* If it's not ours, can't be delete race, (we hold lock). */
 376                 goto no_kprobe;
 377         }
 378
 379         lock_kprobes();
 380         p = get_kprobe(addr);
 381         if (!p) {
 382                 unlock_kprobes();
 383                 if (*addr != BREAKPOINT_INSTRUCTION) {
 384                         /*
 385                          * The breakpoint instruction was removed right
 386                          * after we hit it.  Another cpu has removed
 387                          * either a probepoint or a debugger breakpoint
 388                          * at this address.  In either case, no further
 389                          * handling of this interrupt is appropriate.
 390                          */
 391                         ret = 1;
 392                 }
 393                 /* Not one of ours: let kernel handle it */
 394                 goto no_kprobe;
 395         }
 396
 397         kprobe_status = KPROBE_HIT_ACTIVE;
 398         set_current_kprobe(p, regs);
 399
 400         if (p->pre_handler && p->pre_handler(p, regs))
 401                 /* handler has already set things up, so skip ss setup */
 402                 return 1;
 403
 404 ss_probe:
 405         prepare_singlestep(p, regs);
 406         kprobe_status = KPROBE_HIT_SS;
 407         return 1;
 408
 409 no_kprobe:
 410         preempt_enable_no_resched();
 411         return ret;
 412 }
 413
 414 /*
 415  * For function-return probes, init_kprobes() establishes a probepoint
 416  * here. When a retprobed function returns, this probe is hit and
 417  * trampoline_probe_handler() runs, calling the kretprobe's handler.
 418  */
 419  void kretprobe_trampoline_holder(void)
 420  {
 421         asm volatile (  ".global kretprobe_trampoline\n"
 422                         "kretprobe_trampoline: \n"
 423                         "nop\n");
 424  }
 425
 426 /*
 427  * Called when we hit the probe point at kretprobe_trampoline
 428  */
 429 int trampoline_probe_handler(struct kprobe *p, struct pt_regs *regs)
 430 {
 431         struct task_struct *tsk;
 432         struct kretprobe_instance *ri;
 433         struct hlist_head *head;
 434         struct hlist_node *node;
 435         unsigned long *sara = (unsigned long *)regs->rsp - 1;
 436
 437         tsk = arch_get_kprobe_task(sara);
 438         head = kretprobe_inst_table_head(tsk);
 439
 440         hlist_for_each_entry(ri, node, head, hlist) {
 441                 if (ri->stack_addr == sara && ri->rp) {
 442                         if (ri->rp->handler)
 443                                 ri->rp->handler(ri, regs);
 444                 }
 445         }
 446         return 0;
 447 }
 448
 449 void trampoline_post_handler(struct kprobe *p, struct pt_regs *regs,
 450                                                 unsigned long flags)
 451 {
 452         struct kretprobe_instance *ri;
 453         /* RA already popped */
 454         unsigned long *sara = ((unsigned long *)regs->rsp) - 1;
 455
 456         while ((ri = get_rp_inst(sara))) {
 457                 regs->rip = (unsigned long)ri->ret_addr;
 458                 recycle_rp_inst(ri);
 459         }
 460         regs->eflags &= ~TF_MASK;
 461 }
 462
 463 /*
 464  * Called after single-stepping.  p->addr is the address of the
 465  * instruction whose first byte has been replaced by the "int 3"
 466  * instruction.  To avoid the SMP problems that can occur when we
 467  * temporarily put back the original opcode to single-step, we
 468  * single-stepped a copy of the instruction.  The address of this
 469  * copy is p->ainsn.insn.
 470  *
 471  * This function prepares to return from the post-single-step
 472  * interrupt.  We have to fix up the stack as follows:
 473  *
 474  * 0) Except in the case of absolute or indirect jump or call instructions,
 475  * the new rip is relative to the copied instruction.  We need to make
 476  * it relative to the original instruction.
 477  *
 478  * 1) If the single-stepped instruction was pushfl, then the TF and IF
 479  * flags are set in the just-pushed eflags, and may need to be cleared.
 480  *
 481  * 2) If the single-stepped instruction was a call, the return address
 482  * that is atop the stack is the address following the copied instruction.
 483  * We need to make it the address following the original instruction.
 484  */
 485 static void resume_execution(struct kprobe *p, struct pt_regs *regs)
 486 {
 487         unsigned long *tos = (unsigned long *)regs->rsp;
 488         unsigned long next_rip = 0;
 489         unsigned long copy_rip = (unsigned long)p->ainsn.insn;
 490         unsigned long orig_rip = (unsigned long)p->addr;
 491         kprobe_opcode_t *insn = p->ainsn.insn;
 492
 493         /*skip the REX prefix*/
 494         if (*insn >= 0x40 && *insn <= 0x4f)
 495                 insn++;
 496
 497         switch (*insn) {
 498         case 0x9c:              /* pushfl */
 499                 *tos &= ~(TF_MASK | IF_MASK);
 500                 *tos |= kprobe_old_rflags;
 501                 break;
 502         case 0xc3:              /* ret/lret */
 503         case 0xcb:
 504         case 0xc2:
 505         case 0xca:
 506                 regs->eflags &= ~TF_MASK;
 507                 /* rip is already adjusted, no more changes required*/
 508                 return;
 509         case 0xe8:              /* call relative - Fix return addr */
 510                 *tos = orig_rip + (*tos - copy_rip);
 511                 break;
 512         case 0xff:
 513                 if ((*insn & 0x30) == 0x10) {
 514                         /* call absolute, indirect */
 515                         /* Fix return addr; rip is correct. */
 516                         next_rip = regs->rip;
 517                         *tos = orig_rip + (*tos - copy_rip);
 518                 } else if (((*insn & 0x31) == 0x20) ||  /* jmp near, absolute indirect */
 519                            ((*insn & 0x31) == 0x21)) {  /* jmp far, absolute indirect */
 520                         /* rip is correct. */
 521                         next_rip = regs->rip;
 522                 }
 523                 break;
 524         case 0xea:              /* jmp absolute -- rip is correct */
 525                 next_rip = regs->rip;
 526                 break;
 527         default:
 528                 break;
 529         }
 530
 531         regs->eflags &= ~TF_MASK;
 532         if (next_rip) {
 533                 regs->rip = next_rip;
 534         } else {
 535                 regs->rip = orig_rip + (regs->rip - copy_rip);
 536         }
 537 }
 538
 539 /*
 540  * Interrupts are disabled on entry as trap1 is an interrupt gate and they
 541  * remain disabled thoroughout this function.  And we hold kprobe lock.
 542  */
 543 int post_kprobe_handler(struct pt_regs *regs)
 544 {
 545         if (!kprobe_running())
 546                 return 0;
 547
 548         if ((kprobe_status != KPROBE_REENTER) && current_kprobe->post_handler) {
 549                 kprobe_status = KPROBE_HIT_SSDONE;
 550                 current_kprobe->post_handler(current_kprobe, regs, 0);
 551         }
 552
 553         if (current_kprobe->post_handler != trampoline_post_handler)
 554                 resume_execution(current_kprobe, regs);
 555         regs->eflags |= kprobe_saved_rflags;
 556
 557         /* Restore the original saved kprobes variables and continue. */
 558         if (kprobe_status == KPROBE_REENTER) {
 559                 restore_previous_kprobe();
 560                 goto out;
 561         } else {
 562                 unlock_kprobes();
 563         }
 564 out:
 565         preempt_enable_no_resched();
 566
 567         /*
 568          * if somebody else is singlestepping across a probe point, eflags
 569          * will have TF set, in which case, continue the remaining processing
 570          * of do_debug, as if this is not a probe hit.
 571          */
 572         if (regs->eflags & TF_MASK)
 573                 return 0;
 574
 575         return 1;
 576 }
 577
 578 /* Interrupts disabled, kprobe_lock held. */
 579 int kprobe_fault_handler(struct pt_regs *regs, int trapnr)
 580 {
 581         if (current_kprobe->fault_handler
 582             && current_kprobe->fault_handler(current_kprobe, regs, trapnr))
 583                 return 1;
 584
 585         if (kprobe_status & KPROBE_HIT_SS) {
 586                 resume_execution(current_kprobe, regs);
 587                 regs->eflags |= kprobe_old_rflags;
 588
 589                 unlock_kprobes();
 590                 preempt_enable_no_resched();
 591         }
 592         return 0;
 593 }
 594
 595 /*
 596  * Wrapper routine for handling exceptions.
 597  */
 598 int kprobe_exceptions_notify(struct notifier_block *self, unsigned long val,
 599                              void *data)
 600 {
 601         struct die_args *args = (struct die_args *)data;
 602         switch (val) {
 603         case DIE_INT3:
 604                 if (kprobe_handler(args->regs))
 605                         return NOTIFY_STOP;
 606                 break;
 607         case DIE_DEBUG:
 608                 if (post_kprobe_handler(args->regs))
 609                         return NOTIFY_STOP;
 610                 break;
 611         case DIE_GPF:
 612                 if (kprobe_running() &&
 613                     kprobe_fault_handler(args->regs, args->trapnr))
 614                         return NOTIFY_STOP;
 615                 break;
 616         case DIE_PAGE_FAULT:
 617                 if (kprobe_running() &&
 618                     kprobe_fault_handler(args->regs, args->trapnr))
 619                         return NOTIFY_STOP;
 620                 break;
 621         default:
 622                 break;
 623         }
 624         return NOTIFY_DONE;
 625 }
 626
 627 int setjmp_pre_handler(struct kprobe *p, struct pt_regs *regs)
 628 {
 629         struct jprobe *jp = container_of(p, struct jprobe, kp);
 630         unsigned long addr;
 631
 632         jprobe_saved_regs = *regs;
 633         jprobe_saved_rsp = (long *) regs->rsp;
 634         addr = (unsigned long)jprobe_saved_rsp;
 635         /*
 636          * As Linus pointed out, gcc assumes that the callee
 637          * owns the argument space and could overwrite it, e.g.
 638          * tailcall optimization. So, to be absolutely safe
 639          * we also save and restore enough stack bytes to cover
 640          * the argument area.
 641          */
 642         memcpy(jprobes_stack, (kprobe_opcode_t *) addr, MIN_STACK_SIZE(addr));
 643         regs->eflags &= ~IF_MASK;
 644         regs->rip = (unsigned long)(jp->entry);
 645         return 1;
 646 }
 647
 648 void jprobe_return(void)
 649 {
 650         preempt_enable_no_resched();
 651         asm volatile ("       xchg   %%rbx,%%rsp     \n"
 652                       "       int3                      \n"
 653                       "       .globl jprobe_return_end  \n"
 654                       "       jprobe_return_end:        \n"
 655                       "       nop                       \n"::"b"
 656                       (jprobe_saved_rsp):"memory");
 657 }
 658
 659 int longjmp_break_handler(struct kprobe *p, struct pt_regs *regs)
 660 {
 661         u8 *addr = (u8 *) (regs->rip - 1);
 662         unsigned long stack_addr = (unsigned long)jprobe_saved_rsp;
 663         struct jprobe *jp = container_of(p, struct jprobe, kp);
 664
 665         if ((addr > (u8 *) jprobe_return) && (addr < (u8 *) jprobe_return_end)) {
 666                 if ((long *)regs->rsp != jprobe_saved_rsp) {
 667                         struct pt_regs *saved_regs =
 668                             container_of(jprobe_saved_rsp, struct pt_regs, rsp);
 669                         printk("current rsp %p does not match saved rsp %p\n",
 670                                (long *)regs->rsp, jprobe_saved_rsp);
 671                         printk("Saved registers for jprobe %p\n", jp);
 672                         show_registers(saved_regs);
 673                         printk("Current registers\n");
 674                         show_registers(regs);
 675                         BUG();
 676                 }
 677                 *regs = jprobe_saved_regs;
 678                 memcpy((kprobe_opcode_t *) stack_addr, jprobes_stack,
 679                        MIN_STACK_SIZE(stack_addr));
 680                 return 1;
 681         }
 682         return 0;
 683 }
 684
 685 /*
 686  * kprobe->ainsn.insn points to the copy of the instruction to be single-stepped.
 687  * By default on x86_64, pages we get from kmalloc or vmalloc are not
 688  * executable.  Single-stepping an instruction on such a page yields an
 689  * oops.  So instead of storing the instruction copies in their respective
 690  * kprobe objects, we allocate a page, map it executable, and store all the
 691  * instruction copies there.  (We can allocate additional pages if somebody
 692  * inserts a huge number of probes.)  Each page can hold up to INSNS_PER_PAGE
 693  * instruction slots, each of which is MAX_INSN_SIZE*sizeof(kprobe_opcode_t)
 694  * bytes.
 695  */
 696 #define INSNS_PER_PAGE (PAGE_SIZE/(MAX_INSN_SIZE*sizeof(kprobe_opcode_t)))
 697 struct kprobe_insn_page {
 698         struct hlist_node hlist;
 699         kprobe_opcode_t *insns;         /* page of instruction slots */
 700         char slot_used[INSNS_PER_PAGE];
 701         int nused;
 702 };
 703
 704 static struct hlist_head kprobe_insn_pages;
 705
 706 /**
 707  * get_insn_slot() - Find a slot on an executable page for an instruction.
 708  * We allocate an executable page if there's no room on existing ones.
 709  */
 710 static kprobe_opcode_t *get_insn_slot(void)
 711 {
 712         struct kprobe_insn_page *kip;
 713         struct hlist_node *pos;
 714
 715         hlist_for_each(pos, &kprobe_insn_pages) {
 716                 kip = hlist_entry(pos, struct kprobe_insn_page, hlist);
 717                 if (kip->nused < INSNS_PER_PAGE) {
 718                         int i;
 719                         for (i = 0; i < INSNS_PER_PAGE; i++) {
 720                                 if (!kip->slot_used[i]) {
 721                                         kip->slot_used[i] = 1;
 722                                         kip->nused++;
 723                                         return kip->insns + (i*MAX_INSN_SIZE);
 724                                 }
 725                         }
 726                         /* Surprise!  No unused slots.  Fix kip->nused. */
 727                         kip->nused = INSNS_PER_PAGE;
 728                 }
 729         }
 730
 731         /* All out of space.  Need to allocate a new page. Use slot 0.*/
 732         kip = kmalloc(sizeof(struct kprobe_insn_page), GFP_KERNEL);
 733         if (!kip) {
 734                 return NULL;
 735         }
 736
 737         /*
 738          * For the %rip-relative displacement fixups to be doable, we
 739          * need our instruction copy to be within +/- 2GB of any data it
 740          * might access via %rip.  That is, within 2GB of where the
 741          * kernel image and loaded module images reside.  So we allocate
 742          * a page in the module loading area.
 743          */
 744         kip->insns = module_alloc(PAGE_SIZE);
 745         if (!kip->insns) {
 746                 kfree(kip);
 747                 return NULL;
 748         }
 749         INIT_HLIST_NODE(&kip->hlist);
 750         hlist_add_head(&kip->hlist, &kprobe_insn_pages);
 751         memset(kip->slot_used, 0, INSNS_PER_PAGE);
 752         kip->slot_used[0] = 1;
 753         kip->nused = 1;
 754         return kip->insns;
 755 }
 756
 757 /**
 758  * free_insn_slot() - Free instruction slot obtained from get_insn_slot().
 759  */
 760 static void free_insn_slot(kprobe_opcode_t *slot)
 761 {
 762         struct kprobe_insn_page *kip;
 763         struct hlist_node *pos;
 764
 765         hlist_for_each(pos, &kprobe_insn_pages) {
 766                 kip = hlist_entry(pos, struct kprobe_insn_page, hlist);
 767                 if (kip->insns <= slot
 768                     && slot < kip->insns+(INSNS_PER_PAGE*MAX_INSN_SIZE)) {
 769                         int i = (slot - kip->insns) / MAX_INSN_SIZE;
 770                         kip->slot_used[i] = 0;
 771                         kip->nused--;
 772                         if (kip->nused == 0) {
 773                                 /*
 774                                  * Page is no longer in use.  Free it unless
 775                                  * it's the last one.  We keep the last one
 776                                  * so as not to have to set it up again the
 777                                  * next time somebody inserts a probe.
 778                                  */
 779                                 hlist_del(&kip->hlist);
 780                                 if (hlist_empty(&kprobe_insn_pages)) {
 781                                         INIT_HLIST_NODE(&kip->hlist);
 782                                         hlist_add_head(&kip->hlist,
 783                                                 &kprobe_insn_pages);
 784                                 } else {
 785                                         module_free(NULL, kip->insns);
 786                                         kfree(kip);
 787                                 }
 788                         }
 789                         return;
 790                 }
 791         }
 792 }